Pith. sign in

REVIEW 5 major objections 5 minor 3 cited by

ExEBench: Benchmarking Foundation Models on Extreme Earth Events

T0 review · 5 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A public benchmark scores AI across seven extreme weather disasters

desk verdict A useful multi-event benchmark that is undermined by country-scale heatwave definitions and a few internal inconsistencies; worth reviewing but needs revision. read the letter →

arxiv 2505.08529 v1 pith:YYA2Q6B3 submitted 2025-05-13 cs.LG cs.AI

Shan Zhao , Zhitong Xiong , Jie Zhao , Xiao Xiang Zhu This is my paper · ORCID
classification cs.LGcs.AI
keywords extremeeventsfoundationmodelsbenchmarkEarthobservationweatherandclimatedisastermanagementnowcastingremotesensing
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

ExEBench argues that foundation-model reliability for disaster management should be judged on a common testbed spanning seven extreme-event categories—floods, wildfires, storms, tropical cyclones, extreme precipitation, heatwaves, and cold waves—rather than on the single-domain benchmarks used previously. The paper assembles global datasets from satellite, radar, and reanalysis sources, defines one machine-learning task per event type (detection, monitoring, or forecasting), and reports baselines across vision, Earth-observation, and weather-climate foundation models. Its central message is that pre-trained weights transfer best when the pretraining data and the downstream event share spectral, spatial, and temporal characteristics, while temporal dynamics, extreme-value tails, and sensor-specific cues such as radar coherence remain weak points. Thus the contribution is a reusable public platform plus an initial map of where current foundation models generalize and where they fail on the rare, high-impact events that climate change is making more frequent.

What carries the argument

The load-bearing object is the benchmark itself: seven per-event datasets with fixed train/test splits and a task definition that exposes a different capability—heatwave and cold-wave single-step forecasting, cyclone multi-variable forecasting, storm and extreme-precipitation nowcasting, fire segmentation, and flood change detection. What carries the argument is the controlled pairing of heterogeneous data (coarse global reanalysis fields, high-resolution radar, multispectral imagery, and synthetic-aperture radar) with per-task evaluation metrics, so that model differences can be attributed to pretraining domain, fine-tuning strategy, and architectural choices. The same event-recognition logic runs through every dataset: identify an extreme event from documented or statistical evidence, crop the relevant Earth-observation or weather variables around it, and frame an operational prediction problem.

What would settle it

Re-extract the same 2019–2023 heatwaves using a gridded definition (for example, at least three consecutive days above a local 95th-percentile temperature threshold) and rerun the 10-day forecast evaluation; if ACC or RMSE changes materially—say accuracy drops below the reported range for the leading model—the country-polygon localization assumption is load-bearing.

Watch

Extended reading notes

Core claim

The paper's central claim is that ExEBench offers a fair, reproducible way to test how foundation models behave on extreme Earth events, with seven event categories spanning weather and Earth observation and tasks chosen to mirror operational demands. Each dataset is constructed to preserve real-world heterogeneity: heatwaves and cold waves are extracted from documented disasters and matched to reanalysis temperature fields, extreme precipitation is defined by local 95th-percentile thresholds from a 22-year climatology, tropical cyclones are bounded by observed tracks, storms come from radar sequences, and fires and floods retain their original multiband imagery and masks. The baselines then compare the same model families under frozen, fully fine-tuned, and parameter-efficient settings. The empirical finding that carries the argument is asymmetric transfer: models generalize across modalities only to a point, perform best when pretraining and downstream data share a domain, and degrade sharply on temporal forecasting and on imbalanced decision classes such as urban flooding.

Load-bearing premise

That the documented disasters' country-level polygons closely mark the areas actually affected by each heatwave or cold wave, so extracting temperature over the entire country yields event-relevant training data.

Editorial extensions

If this is right

  • A team deploying a foundation model for disaster response can benchmark it on the public testbed against fixed baseline scores, turning qualitative claims like 'the model handles floods' into measurable comparisons.
  • If the reported transfer pattern holds, model selection should be driven by domain alignment: weather-climate models for forecasting tasks, Earth-observation models for image segmentation, with vision models as a middle ground for imagery.
  • The temporal-dynamics failures set a design target: future foundation models need explicit temporal processing rather than treating time as extra channels to handle storm nowcasting and cyclone evolution.
  • The flood results imply that sensor-aware features—polarization mode, coherence, pre/co/post-event change—must be encoded explicitly in geospatial foundation models.
  • Parameter-efficient fine-tuning such as LoRA offers a practical accuracy-versus-cost compromise when fully fine-tuning is too expensive, with measurable losses mostly on precipitation extremes.

Reading between the lines

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

  • The country-polygon definition of heatwave and cold-wave events means the forecasting tasks may reward models that reproduce broad regional temperature fields rather than the local anomaly itself; an event-mask variant would shift the benchmark from regional forecast skill to event-localized skill.
  • A natural extension is to study compound and cascading events—for instance, a heatwave priming wildfire conditions or extreme precipitation following a cyclone—because the same reanalysis variables appear across multiple ExEBench tasks.
  • The benchmark currently compares foundation models against each other and against random initialization, not against physical or operational forecast baselines, so the practical added value of foundation models over classical weather prediction remains an open question.
  • A testable follow-up would rerun the precipitation nowcasting tasks with a dedicated extreme-loss objective (for example, weighting high-rain-rate bins) to see whether the documented underestimation of extremes is primarily a data-imbalance or an architecture problem.
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

5 major / 5 minor

Summary. The paper introduces ExEBench, a benchmark of seven extreme-event categories (floods, wildfires, storms, tropical cyclones, extreme precipitation, heatwaves, cold waves) built from heterogeneous EO and weather/climate data. For each category it defines an ML task (forecasting, nowcasting, segmentation, change detection), provides standardized train/test splits, and evaluates a suite of foundation models under several fine-tuning strategies. The headline claims are that the benchmark supports assessment of FM transferability across domains, that the reported baselines provide initial evidence on cross-data-type generalization, and that the released public dataset and code constitute a reusable resource for disaster-management ML.

Significance. If the benchmark construction is valid, ExEBench fills a real gap: most existing benchmarks are single-domain, whereas extreme events naturally span EO and weather/climate modalities. The paper's strengths are its public release of data and code, the detailed per-event data-processing workflows (Section 2, Appendix 6.1), the broad model coverage (vision, EO, and W&C foundation models), and the use of domain-appropriate metrics (ACC, RQE, POD/FAR/CSI/HSS, IoU/F1). The reported experiments also give concrete, falsifiable evidence that pre-training domain alignment matters, e.g., Aurora's strong heatwave and precipitation results and ClimaX's poor fire-mapping results. However, the central claim that the tasks are aligned with operational extreme-event needs is currently undercut by the country-level definition of heatwave/cold-wave regions and by an inconsistency in the ACC definition, so the significance of the benchmark as stated is contingent on those issues being resolved.

major comments (5)
  1. [Section 2.2, Figure 2; Tables 3 and 4] The heatwave and cold-wave ROI is defined by the country code (ISO-3) from EM-DAT, and ERA5 t2m is then extracted over the entire country. This conflates "an extreme event occurred somewhere in this country" with "the whole country is the event region." Because large-area temperature fields are spatially smooth and climatologically predictable, the high ACC values (e.g., Aurora ACC 0.8887 for heatwaves at 10 days, Table 3) may largely reflect skill at reproducing broad climatological gradients rather than event-localized extreme-temperature skill. A direct ablation is needed: compare the reported models against a climatology or persistence baseline over the same country-scale regions, and also evaluate on event-localized polygons derived from a meteorological heatwave definition (e.g., EHF or percentile-based contiguous regions). Without such validation, the claim that these tasks measure extreme-event prediction rather than general temperature forecasting is not established.
  2. [Section 6.2.1, Eq. (5)] The ACC formula as written is not the standard anomaly correlation coefficient: the numerator uses (xhat_i - xbar_hat)(x_i - xbar), where xbar_hat is the mean of the forecast field and xbar is the climatology, but a proper ACC should subtract the same climatological reference from both forecast and observation. The text states that xbar is the climatology, which makes the first factor inconsistent. As written, the ACC values in Tables 3-6 are not reproducible, and if the code implements the formula literally, the reported skill scores are not anomaly correlations. Please correct Eq. (5) and, if the experiments used the standard definition, state this explicitly in the text.
  3. [Table 2 and Sections 6.3, 6.3.3, 6.4.4, 6.4.5] There are internal inconsistencies about which models were actually evaluated. Table 2 lists U-Net as pretrained on LGG Brain MRI, but the table note and Section 6.3 state that U-Net is randomly initialized and trained end-to-end. More importantly, Table 2 marks SegFormer and ConvNeXt as evaluated on storms and U-Net and Prithvi as evaluated on extreme precipitation, while Sections 6.3.3, 6.4.4, and 6.4.5 report only Prithvi and Aurora for these two tasks and state that U-Net, SegFormer, and ConvNeXt failed to produce useful outputs and were excluded. Please reconcile Table 2 with the actual experiment log; a benchmark paper should have a fully accurate experiment summary table.
  4. [Section 2.5, Figure 3, and Table 1] The resolution of the extreme-precipitation dataset is described inconsistently. Section 2.5 says the final sequences have a 50x50 spatial size at 0.1-degree resolution, but Figure 3 and Appendix 6.1 describe coarsening both TRMM and IMERG to 5 degrees for threshold computation, and Figure 3's final panel labels the data as "IMERG (5 degree)." Table 1 also lists 0.1-degree resolution. Please clarify whether the 50x50 tiles are local high-resolution tiles around events (with thresholds computed at 5 degrees) or whether the entire dataset is at 5 degrees; this is essential for reproducing the dataset and interpreting the reported spatial-resolution transfer results.
  5. [Sections 3, 6.4, and 4] The interpretation of the baseline results is limited by the absence of any non-FM reference models. For forecasting tasks, simple baselines such as persistence and climatology, and for segmentation tasks a simple texture/spectral baseline, would establish what fraction of the reported skill is attributable to FM transferability rather than to the inherent predictability of the target field. The paper itself acknowledges this in Section 4 ("incorporating baseline results from traditional physical or task-specific models would allow for a fair understanding"), but the absence of such baselines in the main results is a load-bearing gap for the paper's claim that FMs show meaningful generalizability on these tasks.
minor comments (5)
  1. [Section 2.8 vs. Appendix 6.1] The flood dataset description in Section 2.8 lists "VV and VH intensities from pre- and post-event, as well as pre- and co-event coherence," whereas Appendix 6.1 lists eight bands as pre-event intensity, post-event intensity, pre-event coherence, and co-event intensity (with no co-event coherence). Please make these two descriptions consistent.
  2. [Table 2] The footnote explaining red and green checkmarks is difficult to parse; the phrase "U-Net is randomly initialized" belongs in the model column or in the caption rather than as a color-code explanation, and it should be consistent with the "Pretrained data" column entry (Brain MRI).
  3. [Section 6.2.1] Equation (5) also contains typographical artifacts (extra parentheses and a missing superscript on the first denominator term); please re-typeset the equation and verify all symbols against the standard ACC definition.
  4. [Table 1] The evaluation metric for tropical cyclones is listed as "Trajectory tracking," but the defined task in Section 2.6 is multi-variable weather forecasting (Eq. 3) and the reported metrics are RMSE and ACC. Please align the table's task description with the actual task.
  5. [Section 4] The sentence "for the first time, tested cross-data-type generalization" is a strong claim that would benefit from a precise statement of what is new relative to prior transfer experiments between EO and weather/climate models, or from softening the wording.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: ExEBench is a dataset and benchmark paper whose claims are supported by external data sources and measured model outputs, not by self-referential derivations.

full rationale

ExEBench is a benchmark construction paper rather than a derivation of a scientific quantity from fitted inputs. The central claims are that the benchmark integrates seven extreme-event datasets from external sources (EM-DAT, ERA5, TRMM/IMERG, IBTrACS, TASSRAD19, HLS burn scars, UrbanSARFloods) and that the defined ML tasks evaluate foundation models with reported metrics. No quantity is fitted to a subset of data and then renamed as a prediction; the ACC, RMSE, RQE, POD, CSI, HSS, F1, and IoU scores are all computed by comparing model outputs against independent ground-truth targets. The paper's self-citations (e.g., DOFA, UrbanSARFloods, prior precipitation-nowcasting work by the same authors) are used as dataset sources, baseline models, or standard references, and none of them is invoked to justify a uniqueness theorem, to force a choice of ansatz, or to establish the benchmark's validity by author fiat. The identified concern about heatwave/cold-wave event regions being approximated by whole-country ISO-3 codes from EM-DAT is a data-validity or task-design limitation, not circularity: the forecasting target is still ERA5 temperature over that region, and the prediction is not equivalent to the event-selection input by construction. The benchmark is therefore self-contained with respect to external benchmarks and data sources, and no circular step meets the evidence threshold requiring a quoted reduction.

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

All parameters and assumptions shape the dataset. The benchmark's validity rests on the external data sources being accurate enough for their intended role, and on the paper's event-selection thresholds matching the operational notion of 'extreme event'. No new physical entities are introduced.

free parameters (5)
  • Extreme precipitation 95th percentile threshold (RP95) = Climatological 95th percentile per 5-degree grid cell and month, derived from TRMM 1998-2019
    Defines which IMERG periods count as extreme precipitation events; the paper supplies this threshold rather than an external standard (Section 2.5, Appx. 6.1).
  • Consecutive-days criterion for extreme precipitation events = 3 days
    Only periods longer than three days above RP95 are selected; changing this changes the event set (Section 2.5).
  • Heatwave and cold wave duration and buffer settings = Max duration 6 months; 1-month pre-event buffer
    Used to extract HW and CW sequences from EM-DAT; no external standard is cited for these choices (Sections 2.2, 2.3).
  • Tropical cyclone maximum duration = 14 days
    Filters IBTrACS events; affects dataset composition and task difficulty (Section 2.6).
  • Storm keyword filter = Keywords: storm, downpour, hail
    Selects radar sequences from TASSRAD19 based on daily weather summaries; the keyword set is subjective (Section 2.4).
assumptions (5)
  • domain assumption EM-DAT records accurately identify extreme heat and cold events, including their country-level spatial extent and start and end dates.
    Section 2.2 relies on country codes and EM-DAT dates to define event ROIs; if EM-DAT is incomplete or spatially coarse, the tasks are mis-specified.
  • domain assumption ERA5 reanalysis fields are a faithful proxy for observed weather in heatwave, cold wave, and tropical cyclone forecasting.
    HW, CW, and TC targets are ERA5 fields; model skill is measured against reanalysis, not station observations (Sections 2.2, 2.3, 2.6).
  • domain assumption TRMM and IMERG precipitation products are consistent enough after unit and grid conversions to identify extreme events across 1998-2023.
    Section 2.5 and Appx. 6.1 compute thresholds from TRMM and apply them to IMERG; discrepancies between the two products could bias event selection.
  • domain assumption The 95th percentile of daily precipitation is a valid operational definition of extreme precipitation.
    This is a common climatological convention, but it is still a modeling choice; the paper does not validate it against socioeconomic impact data (Section 2.5).
  • domain assumption Ground-truth masks in HLS BurnScars and UrbanSARFloods are accurate labels for burned areas and flooded areas.
    Fire and flood tasks are supervised by these existing benchmark masks; any systematic label errors propagate to model evaluation (Sections 2.7, 2.8).

how reviews work

0 comments
Cite this review

Pith. "Pith review of ExEBench: Benchmarking Foundation Models on Extreme Earth Events." pith.science (2026). https://pith.science/paper/YYA2Q6B3

@misc{pith2026250508529,
  author       = {Pith},
  title        = {Pith review of: ExEBench: Benchmarking Foundation Models on Extreme Earth Events},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YYA2Q6B3}},
  note         = {Machine review of arXiv:2505.08529}
}
read the original abstract

Our planet is facing increasingly frequent extreme events, which pose major risks to human lives and ecosystems. Recent advances in machine learning (ML), especially with foundation models (FMs) trained on extensive datasets, excel in extracting features and show promise in disaster management. Nevertheless, these models often inherit biases from training data, challenging their performance over extreme values. To explore the reliability of FM in the context of extreme events, we introduce \textbf{ExE}Bench (\textbf{Ex}treme \textbf{E}arth Benchmark), a collection of seven extreme event categories across floods, wildfires, storms, tropical cyclones, extreme precipitation, heatwaves, and cold waves. The dataset features global coverage, varying data volumes, and diverse data sources with different spatial, temporal, and spectral characteristics. To broaden the real-world impact of FMs, we include multiple challenging ML tasks that are closely aligned with operational needs in extreme events detection, monitoring, and forecasting. ExEBench aims to (1) assess FM generalizability across diverse, high-impact tasks and domains, (2) promote the development of novel ML methods that benefit disaster management, and (3) offer a platform for analyzing the interactions and cascading effects of extreme events to advance our understanding of Earth system, especially under the climate change expected in the decades to come. The dataset and code are public https://github.com/zhaoshan2/EarthExtreme-Bench.

Figures

Figures reproduced from arXiv: 2505.08529 by the authors.

Figure 1
Figure 1. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Workflow of heatwave extraction. We first identify heatwave events from EmDat database, and then prepare the corresponding local ERA5 variable during the events. 2.4 Storms Precipitation is one of the most important meteorological variables. Its prediction is challenging due to its inherent stochastic nature, involvement of multi-scale processes, and complex patterns [21]. Most DL-based models fail to capture the in… view at source ↗
Figure 3
Figure 3. Workflow of extreme precipitation extraction. Global rainfall events exceeding the local 95th percentile values over the past 22 years are selected. (pcp: precipitation) 2.6 Tropical cyclones Tropical cyclone (TC) is the organized system of thunderstorms that develops over tropical or subtrop￾ical waters. Its formation and development is fast and involves complex and multiple meteorological variables anomalies. Rece… view at source ↗
Figures from the paper (20 more)
Figure 4
Figure 4. Figure 4: Workflow of Tropical cyclone extraction. Tropical cyclone events are selected, and the corresponding meteorological condition data during these events are prepared (a,b). Socioeconomic losses are estimated from the EmDat database (c). six channels. The masks have three…
Figure 5
Figure 5. Figure 5: Workflow of Storms: Radar sequences are filtered based on daily weather summaries con￾taining keywords such as “storm”. b. Select the data from the flooded urban areas Non-flooded areas Flooded urban areas Flooded open areas a. Download UrbanSARFloods 8 bands of Pre-pr…
Figure 7
Figure 7. Figure 7: Datasets in our benchmark exhibit diverse characteristics across multiple dimensions. As the radial axis extends outward, the resolution increases or the quantity becomes higher. Task The task is to segment the input X ∈ R 6×512×512 to three classes Yˆ ∈ R 3×512×512 . …
Figure 8
Figure 8. Figure 8: Sample results. (a) Model performance on normalized scores. A score of 0 denotes either failure to process its data modality or complete task failure. (b) The fire segmentation using pre-trained weights (△) and randomly initialized weights (◦). (c) Fire segmentation un…
Figure 9
Figure 9. Figure 9: Comparison of statistics between a 39-year long-term climatology and extreme events (Heatwaves, cold waves, and tropical cyclones). The mean values observed during extreme events generally show a significant deviation from the long-term climatological averages. However…
Figure 10
Figure 10. Figure 10: Extreme precipitation: The global monthly 95 percentile precipitation rate (mm/d) from 22-year TRMM dataset. Every grid point has its own extreme precipitation threshold, depending on the climatology. 2021 2022 2023 [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Extreme precipitation: The locations of events in the final Expcp dataset. independence, with training data covering flood events in Beira, Iran, Canada, Japan, Port Macquarie, Sydney, Coraki Niger, Hebei, Beledweyne, and the test data covers events from Houston and L…
Figure 12
Figure 12. Figure 12: visualizes the label distribution in the Fire and Flood datasets. Both exhibit label imbalance, a common challenge in many geospatial tasks. (a) Fire (b) Flood [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: Heatwave: 80 to 98 quantile errors of different models for forecasting window of 10 days in advance. (a) U-Net (b) SegFormer (c) ConvNeXt (d) SatMAE (e) Prithvi (f) DOFA* (g) Aurora [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]
Figure 14
Figure 14. Figure 14: Heatwave: Histogram of the labels and predictions of different models for forecasting window of 10 days in advance. (a) U-Net (b) SegFormer (c) ConvNeXt (d) SatMAE (e) Prithvi (f) DOFA* (g) Aurora K K [PITH_FULL_IMAGE:figures/full_fig_p020_14.png]
Figure 15
Figure 15. Figure 15: Heatwave: spatial bias of model outputs, computed as the difference between predictions and ground truth at a 10-day lead time. 6.4.2 Cold waves Results [PITH_FULL_IMAGE:figures/full_fig_p020_15.png]
Figure 16
Figure 16. Figure 16: Heatwave: Visual sample images of model (Aurora) output for forecasting window of 10 days in advance in April to September. could be attributed to U-Net’s relatively smaller number of parameters, which makes it less prone to overfitting in the smaller cold waves datas…
Figure 17
Figure 17. Figure 17: Cold waves: 80 to 98 quantile errors of different models for forecasting window of 5 days in advance. (a) U-Net (b) SegFormer (c) ConvNeXt (d) SatMAE (e) Prithvi (f) DOFA (g) Aurora [PITH_FULL_IMAGE:figures/full_fig_p022_17.png]
Figure 18
Figure 18. Figure 18: Cold waves: Histogram of the labels and predictions of different models for forecasting window of 5 days in advance. Discussion The experiments on tropical cyclones demonstrate the effectiveness of FMs in fine￾tuning from global forecasts to local phenomena while hand…
Figure 19
Figure 19. Figure 19: Tropical cyclones: The visual sample frames for the forecasting window of 6 hours in advance [PITH_FULL_IMAGE:figures/full_fig_p023_19.png]
Figure 20
Figure 20. Figure 20: Storms: The visual comparison of sample frames between different models (Prithvi and Aurora) for a nowcasting window of 2 hours in advance. struggles with the large gap in spatial granularity (from 20 meters to thousands of kilometers, and from regional to global scal…
Figure 21
Figure 21. Figure 21: Extreme precipitation: The visual comparison of sample frames between Aurora and Prithvi for the nowcasting window of 30 minutes in advance. 6.4.6 Fires Results In [PITH_FULL_IMAGE:figures/full_fig_p025_21.png]
Figure 22
Figure 22. Figure 22: Fire: Training loss over the first 10 epochs for different models, fine-tuned with either random initialization (dashed line and marked with “-r”) or pretrained weights. To ensure comparability, the same learning rate was used across all models. Models with pretrained…
Figure 23
Figure 23. Figure 23: Fire prediction maps. We visualize the input using SWIR2 (B12), NIR (B8A), and Red (B04) bands. The yellow pixels in the masks are burned areas [PITH_FULL_IMAGE:figures/full_fig_p026_23.png]
Figure 24
Figure 24. Figure 24 [PITH_FULL_IMAGE:figures/full_fig_p027_24.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Physics-Informed Super-Resolution of Atmospheric Data

    cs.LG 2026-07 reject novelty 5.0 of 10

    Adding multi-scale hydrostatic-primitive-equation losses to atmospheric super-resolution models improves reported physical-consistency scores and some reconstruction/event-detection metrics, but the metric and constra...

  2. The Rise of AI in Weather and Climate Information and its Impact on Global Inequality

    physics.ao-ph 2026-03 conditional novelty 4.0 of 10

    AI weather and climate tools inherit Northern-controlled data and compute, risking worse forecasts and maladaptation for the Global South rather than democratizing climate information.

  3. Scalable and Trustworthy Earth Observation Foundation Models

    cs.LG 2026-07 conditional novelty 3.0 of 10

    Remote-sensing foundation models need domain-specific design and evaluation around measurement physics and decision constraints; benchmark accuracy alone is insufficient for trustworthy EO deployment.

Reference graph

Works this paper leans on

47 extracted references · 39 canonical work pages · cited by 3 Pith papers

  1. [1]

    Weather and climate extreme events in a changing climate,

    Sonia I Seneviratne, Xuebin Zhang, Muhammad Adnan, Wafae Badi, Claudine Dereczynski, A Di Luca, Subimal Ghosh, Iskhaq Iskandar, James Kossin, Sophie Lewis, et al., “Weather and climate extreme events in a changing climate,” 2021

    work page 2021

  2. [2]

    thesis, GIEC, 2022

    Camille Parmesan, Mike D Morecroft, and Yongyut Trisurat,Climate change 2022: Impacts, adaptation and vulnerability, Ph.D. thesis, GIEC, 2022

    work page 2022

  3. [3]

    Antici- pating cascading effects of extreme precipitation with pathway schemes-three case studies from europe,

    Simone Schauwecker, Estíbaliz Gascón, Shinju Park, Virginia Ruiz-Villanueva, Manfred Schwarb, Daniel Sempere-Torres, Markus Stoffel, Claudia Vitolo, and Mario Rohrer, “Antici- pating cascading effects of extreme precipitation with pathway schemes-three case studies from europe,”Environment international, vol. 127, pp. 291–304, 2019

    work page 2019

  4. [4]

    Crimmins, Justin M

    Deepti Singh, Allison R. Crimmins, Justin M. Pflug, Patrick L. Barnard, Jennifer F. Helgeson, Andrew Hoell, Fayola H. Jacobs, Michael G. Jacox, Alessandra Jerolleman, and Michael F. Wehner,Focus on compound events, U.S. Global Change Research Program, Washington, DC, USA, 2023

    work page 2023

  5. [5]

    On the foundations of earth and climate foundation models,

    Xiao Xiang Zhu, Zhitong Xiong, Yi Wang, Adam J Stewart, Konrad Heidler, Yuanyuan Wang, Zhenghang Yuan, Thomas Dujardin, Qingsong Xu, and Yilei Shi, “On the foundations of earth and climate foundation models,”arXiv preprint arXiv:2405.04285, 2024

    arXiv 2024

  6. [6]

    Neural plasticity- inspired foundation model for observing the earth crossing modalities,

    Zhitong Xiong, Yi Wang, Fahong Zhang, Adam J Stewart, Joëlle Hanna, Damian Borth, Ioannis Papoutsis, Bertrand Le Saux, Gustau Camps-Valls, and Xiao Xiang Zhu, “Neural plasticity- inspired foundation model for observing the earth crossing modalities,”arXiv e-prints, pp. arXiv–2403, 2024

    work page 2024

  7. [7]

    Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery,

    Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon, “Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery,”Advances in Neural Information Processing Systems, vol. 35, pp. 197–211, 2022

    work page 2022

  8. [8]

    Prithvi wxc: Foundation model for weather and climate,

    Johannes Schmude, Sujit Roy, Will Trojak, Johannes Jakubik, Daniel Salles Civitarese, Shraddha Singh, Julian Kuehnert, Kumar Ankur, Aman Gupta, Christopher E Phillips, et al., “Prithvi wxc: Foundation model for weather and climate,”arXiv preprint arXiv:2409.13598, 2024

    arXiv 2024

Show all 47 references
  1. [9]

    Climax: A foundation model for weather and climate,

    Tung Nguyen, Johannes Brandstetter, Ashish Kapoor, Jayesh K Gupta, and Aditya Grover, “Climax: A foundation model for weather and climate,” inInternational Conference on Machine Learning. PMLR, 2023, pp. 25904–25938

    2023

  2. [10]

    What if? numerical weather prediction at the crossroads,

    Peter Bauer, “What if? numerical weather prediction at the crossroads,”Journal of the European Meteorological Society, vol. 1, pp. 100002, 2024

    2024

  3. [11]

    SustainBench: Benchmarks for monitoring the sustainable development goals with machine learning,

    Christopher Yeh, Chenlin Meng, Sherrie Wang, Anne Driscoll, Erik Rozi, Patrick Liu, Jihyeon Lee, Marshall Burke, David B Lobell, and Stefano Ermon, “SustainBench: Benchmarks for monitoring the sustainable development goals with machine learning,” inThirty-fifth Conference on N...

  4. [12]

    Geo- bench: Toward foundation models for earth monitoring,

    Alexandre Lacoste, Nils Lehmann, Pau Rodriguez, Evan Sherwin, Hannah Kerner, Björn Lütjens, Jeremy Irvin, David Dao, Hamed Alemohammad, Alexandre Drouin, et al., “Geo- bench: Toward foundation models for earth monitoring,”Advances in Neural Information Processing Systems, vol....

    2024

  5. [13]

    Fomo-bench: a multi-modal, multi-scale and multi-task forest monitoring benchmark for remote sensing foundation models,

    Nikolaos Ioannis Bountos, Arthur Ouaknine, and David Rolnick, “Fomo-bench: a multi-modal, multi-scale and multi-task forest monitoring benchmark for remote sensing foundation models,” arXiv preprint arXiv:2312.10114, 2023

    2023 arXiv

  6. [14]

    Weatherbench: a benchmark data set for data-driven weather forecasting,

    Stephan Rasp, Peter D Dueben, Sebastian Scher, Jonathan A Weyn, Soukayna Mouatadid, and Nils Thuerey, “Weatherbench: a benchmark data set for data-driven weather forecasting,” Journal of Advances in Modeling Earth Systems, vol. 12, no. 11, pp. e2020MS002203, 2020. 11

    2020

  7. [15]

    Weatherbench 2: A benchmark for the next generation of data-driven global weather models,

    Stephan Rasp, Stephan Hoyer, Alexander Merose, Ian Langmore, Peter Battaglia, Tyler Russell, Alvaro Sanchez-Gonzalez, Vivian Yang, Rob Carver, Shreya Agrawal, et al., “Weatherbench 2: A benchmark for the next generation of data-driven global weather models,”Journal of Advances...

    2024

  8. [16]

    Defining extreme events: A cross-disciplinary review,

    Lauren E McPhillips, Heejun Chang, Mikhail V Chester, Yaella Depietri, Erin Friedman, Nancy B Grimm, John S Kominoski, Timon McPhearson, Pablo Méndez-Lázaro, Emma J Rosi, et al., “Defining extreme events: A cross-disciplinary review,”Earth’s Future, vol. 6, no. 3, pp. 441–455, 2018

    2018

  9. [17]

    Heatwave,

    World Meteorological Organization (WMO), “Heatwave,” https://wmo.int/topics/ heatwave, Accessed: 2025-01-07

    2025

  10. [18]

    Heat-related mortality in europe during the summer of 2022,

    Joan Ballester, Marcos Quijal-Zamorano, Raúl Fernando Méndez Turrubiates, Ferran Pegenaute, François R Herrmann, Jean Marie Robine, Xavier Basagaña, Cathryn Tonne, Josep M Antó, and Hicham Achebak, “Heat-related mortality in europe during the summer of 2022,”Nature medicine, v...

    2022

  11. [19]

    EM-DAT: The International Disaster Database,

    CRED/UCLouvain, “EM-DAT: The International Disaster Database,” 2024, Accessed: 2024- 06-17

    2024

  12. [20]

    The ERA5 global reanalysis,

    Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, András Horányi, Joaquín Muñoz- Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, et al., “The ERA5 global reanalysis,”Quarterly Journal of the Royal Meteorological Society, vol. 146, no. 730, pp. 19...

    1999

  13. [21]

    Exploring geo- metric deep learning for precipitation nowcasting,

    Shan Zhao, Sudipan Saha, Zhitong Xiong, Niklas Boers, and Xiao Xiang Zhu, “Exploring geo- metric deep learning for precipitation nowcasting,” inIGARSS 2023-2023 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2023, pp. 3760–3763

    2023

  14. [22]

    Rainnet v1. 0: a convolutional neural network for radar-based precipitation nowcasting,

    Georgy Ayzel, Tobias Scheffer, and Maik Heistermann, “Rainnet v1. 0: a convolutional neural network for radar-based precipitation nowcasting,”Geoscientific Model Development, vol. 13, no. 6, pp. 2631–2644, 2020

    2020

  15. [23]

    TAASRAD19, a high-resolution weather radar reflectivity dataset for precipitation nowcasting,

    Gabriele Franch, Valerio Maggio, Luca Coviello, Marta Pendesini, Giuseppe Jurman, and Cesare Furlanello, “TAASRAD19, a high-resolution weather radar reflectivity dataset for precipitation nowcasting,”Scientific Data, vol. 7, no. 1, pp. 234, 2020

    2020

  16. [24]

    The trmm multisatellite precipitation analysis (tmpa): Quasi-global, multiyear, combined-sensor precipitation estimates at fine scales,

    George J Huffman, David T Bolvin, Eric J Nelkin, David B Wolff, Robert F Adler, Guojun Gu, Yang Hong, Kenneth P Bowman, and Erich F Stocker, “The trmm multisatellite precipitation analysis (tmpa): Quasi-global, multiyear, combined-sensor precipitation estimates at fine scales,...

    2007

  17. [25]

    Complex networks reveal global pattern of extreme-rainfall teleconnections,

    Niklas Boers, Bedartha Goswami, Aljoscha Rheinwalt, Bodo Bookhagen, Brian Hoskins, and Jürgen Kurths, “Complex networks reveal global pattern of extreme-rainfall teleconnections,” Nature, vol. 566, no. 7744, pp. 373–377, 2019

    2019

  18. [26]

    GPM IMERG Final Precipitation L3 Half Hourly 0.1 degree x 0.1 degree V07,

    G.J. Huffman, E.F. Stocker, D.T. Bolvin, E.J. Nelkin, and Jackson Tan, “GPM IMERG Final Precipitation L3 Half Hourly 0.1 degree x 0.1 degree V07,” 2023, Accessed: 11-05-2024

    2023

  19. [27]

    Accurate medium-range global weather forecasting with 3d neural networks,

    Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian, “Accurate medium-range global weather forecasting with 3d neural networks,”Nature, vol. 619, no. 7970, pp. 533–538, 2023

    2023

  20. [28]

    Learning skillful medium-range global weather forecasting,

    Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al., “Learning skillful medium-range global weather forecasting,”Science, vol. 382, no. 6677, pp. 1416–1421, 2023

    2023

  21. [29]

    The international best track archive for climate stewardship (IBTrACS) unifying tropical cyclone data,

    Kenneth R Knapp, Michael C Kruk, David H Levinson, Howard J Diamond, and Charles J Neumann, “The international best track archive for climate stewardship (IBTrACS) unifying tropical cyclone data,”Bulletin of the American Meteorological Society, vol. 91, no. 3, pp. 363–376, 2010. 12

    2010

  22. [30]

    Wildfire danger prediction and understanding with deep learning,

    Spyros Kondylatos, Ioannis Prapas, Michele Ronco, Ioannis Papoutsis, Gustau Camps-Valls, María Piles, Miguel-Ángel Fernández-Torres, and Nuno Carvalhais, “Wildfire danger prediction and understanding with deep learning,”Geophysical Research Letters, vol. 49, no. 17, pp. e2022G...

    2022

  23. [31]

    HLS Foundation Burnscars Dataset,

    Christopher Phillips, Sujit Roy, Kumar Ankur, and Rahul Ramachandran, “HLS Foundation Burnscars Dataset,” Aug. 2023

    2023

  24. [32]

    Urban flood mapping using satellite synthetic aperture radar data: A review of characteristics, approaches, and datasets,

    Jie Zhao, Ming Li, Yu Li, Patrick Matgen, and Marco Chini, “Urban flood mapping using satellite synthetic aperture radar data: A review of characteristics, approaches, and datasets,” IEEE Geoscience and Remote Sensing Magazine, 2024

    2024

  25. [33]

    UrbanSARFloods: Sentinel-1 SLC-based benchmark dataset for urban and open-area flood mapping,

    Jie Zhao, Zhitong Xiong, and Xiao Xiang Zhu, “UrbanSARFloods: Sentinel-1 SLC-based benchmark dataset for urban and open-area flood mapping,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 419–429

    2024

  26. [34]

    Fourcastnet: Acceler- ating global high-resolution weather forecasting using adaptive fourier neural operators,

    Thorsten Kurth, Shashank Subramanian, Peter Harrington, Jaideep Pathak, Morteza Mardani, David Hall, Andrea Miele, Karthik Kashinath, and Anima Anandkumar, “Fourcastnet: Acceler- ating global high-resolution weather forecasting using adaptive fourier neural operators,” in Proc...

    2023

  27. [35]

    U-net: Convolutional networks for biomedical image segmentation,

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-net: Convolutional networks for biomedical image segmentation,” inMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part...

    2015

  28. [36]

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

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo, “Segformer: Simple and efficient design for semantic segmentation with transformers,”Advances in neural information processing systems, vol. 34, pp. 12077–12090, 2021

    2021

  29. [37]

    A convnet for the 2020s,

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie, “A convnet for the 2020s,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 11976–11986

    2022

  30. [38]

    Foundation models for generalist geospatial artificial intelli- gence,

    J Jakubik, S Roy, C Phillips, P Fraccaro, D Godwin, B Zadrozny, D Szwarcman, C Gomes, G Nyirjesy, B Edwards, et al., “Foundation models for generalist geospatial artificial intelli- gence,”arXiv preprint arXiv:2310.18660, 2023

    arXiv

  31. [39]

    Prithvi-eo-2.0: A versatile multi-temporal foundation model for earth observation applications,

    Daniela Szwarcman, Sujit Roy, Paolo Fraccaro, Þorsteinn Elí Gíslason, Benedikt Blumenstiel, Rinki Ghosal, Pedro Henrique de Oliveira, Joao Lucas de Sousa Almeida, Rocco Sedona, Yanghui Kang, et al., “Prithvi-eo-2.0: A versatile multi-temporal foundation model for earth observa...

    2024

  32. [40]

    Aurora: A foundation model for the earth system,

    Cristian Bodnar, Wessel P. Bruinsma, Ana Lucic, Megan Stanley, Anna Vaughan, Johannes Brandstetter, Patrick Garvan, Maik Riechert, Jonathan A. Weyn, Haiyu Dong, Jayesh K. Gupta, Kit Thambiratnam, Alexander T. Archibald, Chun-Chieh Wu, Elizabeth Heider, Max Welling, Richard E. ...

    2024

  33. [41]

    Lora: Low-rank adaptation of large language models,

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen, “Lora: Low-rank adaptation of large language models,”arXiv preprint arXiv:2106.09685, 2021

    2021 arXiv

  34. [42]

    Unified perceptual parsing for scene understanding,

    Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun, “Unified perceptual parsing for scene understanding,” inProceedings of the European conference on computer vision (ECCV), 2018, pp. 418–434

    2018

  35. [43]

    Lgg mri segmentation dataset,

    Mateusz Buda, “Lgg mri segmentation dataset,” https://www.kaggle.com/datasets/ mateuszbuda/lgg-mri-segmentation/data, 2019, Accessed: 2024-11-03

    2019

  36. [44]

    Extreme- cast: Boosting extreme value prediction for global weather forecast,

    Wanghan Xu, Kang Chen, Tao Han, Hao Chen, Wanli Ouyang, and Lei Bai, “Extreme- cast: Boosting extreme value prediction for global weather forecast,”arXiv preprint arXiv:2402.01295, 2024. 13

    2024 arXiv

  37. [45]

    Deep learning for precipitation nowcasting: A benchmark and a new model,

    Xingjian Shi, Zhihan Gao, Leonard Lausen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun Woo, “Deep learning for precipitation nowcasting: A benchmark and a new model,” Advances in neural information processing systems, vol. 30, 2017

    2017

  38. [46]

    Fixing weight decay regularization in adam,

    Ilya Loshchilov, Frank Hutter, et al., “Fixing weight decay regularization in adam,”

  39. [47]

    Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations,

    Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso, “Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations,” inDeep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support:...

    2017

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.