Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

MONITRS: Multimodal Observations of Natural Incidents Through Remote Sensing

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 A new dataset pairing temporal satellite imagery with news-derived text lets multimodal language models detect and track natural disasters, lifting event classification from about 50 percent to 88.7 percent accuracy.

desk verdict A large, novel disaster-monitoring dataset built from FEMA records and news articles, but the headline claim of visual understanding is not yet supported because the text-derived labels were never checked against the imagery. read the letter →

arxiv 2507.16228 v1 pith:WBXP5FNP submitted 2025-07-22 cs.CV

Shreelekha Revankar , Utkarsh Mall , Cheng Perng Phoo , Kavita Bala , Bharath Hariharan This is my paper
classification cs.CV
keywords disastermonitoringsatelliteimagerymultimodaldatasettemporalgroundingeventclassificationvision-languagemodelsnaturallanguageannotations
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

The paper presents MONITRS, a multimodal dataset of more than 10,000 US disaster events that aligns multi-date satellite imagery with natural-language descriptions extracted from news articles. The central claim is that this dataset can teach existing multimodal language models to recognize disasters and track how they evolve, closing a gap where current models perform near chance. After fine-tuning on about one fifth of the training QA pairs, the paper reports event classification accuracy rising from roughly 50 percent to 88.7 percent, and temporal grounding from 11-18 percent to 70.7 percent. The authors argue this establishes a benchmark for machine-learning-assisted disaster response and a recipe for building similar datasets in other event domains.

What carries the argument

The engine of the paper is the MONITRS curation pipeline. For each event in the federal disaster declaration list that has county, state, dates, and a name, the pipeline retrieves up to five news articles via a web search API. An LLM extracts proper-noun locations from the articles, a geocoding service converts those names to coordinates, and the algorithm selects the 5.12 km x 5.12 km satellite patch containing the most named locations for each date. A second LLM pass writes a per-date caption describing what would be visible in the imagery, using only the article text and the date list. These captions and geotagged coordinates seed two kinds of QA items, templated and generated, which make up MONITRS-QA and provide the training and evaluation signal.

What would settle it

Randomly select a set of MONITRS events, blur or mask the second half of each image sequence, and ask a fine-tuned model to answer the temporal-grounding questions; if accuracy remains near the reported 70.7 percent rather than falling toward chance, the model is not using the visual progression and the benchmark is not measuring what it claims.

Watch

Extended reading notes

Core claim

The core discovery is that a data pipeline combining public disaster declarations, geocoded news reports, and optical satellite image sequences can generate supervision that lets a vision-language model perform fine-grained temporal reasoning about disasters. The paper demonstrates this by fine-tuning a temporal earth-observation model on roughly one fifth of the MONITRS-QA training set, then measuring accuracy on held-out events. Event classification jumps from 48-50 percent for three strong baselines to 88.69 percent for the fine-tuned model; temporal grounding rises from 11-18 percent to 70.72 percent; location grounding improves less dramatically, from 13-17 percent to 23.25 percent. The improvements are statistically significant under McNemar's test, and the paper takes this as evidence that the bottleneck for disaster monitoring is data, not model architecture.

Load-bearing premise

The captions and question answers are generated by a language model from news-article text alone, with no verification that the described phenomena are actually visible in the satellite images, so the benchmark's validity rests on the assumption that textual descriptions correspond to optically observable disaster signatures within 10-meter resolution imagery.

Editorial extensions

If this is right

  • Disaster responders could query a fine-tuned model in natural language to determine when an event started, how it spread, and which named roads or towns were affected, reducing reliance on manual expert inspection of imagery.
  • The same pipeline can be reused for any event class that is documented in news articles but lacks visual annotations, such as industrial accidents or large-scale social events, extending the dataset's reach beyond natural disasters.
  • Because the paper shows gains with only a fifth of the training data, scaling to the full dataset and adding temporal embeddings or location masks is likely to push accuracy higher, especially for location grounding.
  • The benchmark provides a standardized test for future multimodal models in remote sensing, forcing them to reason across image sequences rather than single frames.

Reading between the lines

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

  • A decisive validity check would be a human study in which annotators who have not read the articles judge whether each LLM-generated caption matches what is visible in the corresponding satellite image; without such a check, part of the measured accuracy could come from the model exploiting common disaster narratives rather than visual evidence.
  • The dataset's coordinate tiling at 5.12 km and 10-meter pixels means it is best suited to large, slow-moving disasters like wildfires and hurricanes; a natural extension would pair the same news-derived text with higher-resolution commercial imagery or synthetic aperture radar to capture cloud-covered or small-scale events.
  • One could test whether the temporal grounding improvements persist when the image order is shuffled or when individual dates are removed; if accuracy collapses, the model is genuinely reading the sequence, and if not, the caption text is doing the work.
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 introduces MONITRS, a large-scale multimodal dataset of approximately 10,000 FEMA disaster events, pairing temporal Sentinel-2 satellite imagery with natural language captions and QA pairs derived from news articles. The authors fine-tune the TEOChat vision-language model on a subset of the dataset and report large accuracy gains on event classification (88.69%) and temporal grounding (70.72%) compared to baseline models, alongside more modest gains on location grounding and open-ended VQA. The dataset is made publicly available, with code released.

Significance. If the dataset's labels are visually grounded, MONITRS would fill a genuine gap in disaster monitoring resources: it provides temporal satellite imagery with natural language annotations across diverse disaster types, enabling training and evaluation of multimodal models for tasks that existing remote-sensing VLM datasets do not cover. The release of code and data is a concrete contribution. However, the central claim that fine-tuning on MONITRS produces models that perform visual disaster monitoring rests on an assumption that the paper itself acknowledges as unverified: the text-derived labels may not reflect what is actually visible in the Sentinel-2 imagery. The fine-tuning improvements in Table 2 are large and statistically significant as reported, but the validity of the benchmark as a measure of visual understanding is not established. Resolving the grounding question is therefore load-bearing for the paper's contribution.

major comments (3)
  1. [Section 3.1, Figure 3] The captions that drive both the dataset and the QA labels are generated by prompting Gemini with news article text and dates 'using the text alone', with the stated intention of captioning 'independently of what may be visible'. Consequently, the QA labels—event classification, temporal grounding, and location grounding—are all derived from text that never referenced the satellite imagery. The paper's Limitations section concedes that 'the descriptions generated by LLMs based on news articles may not always accurately reflect what is visible in the satellite imagery.' Given this, the fine-tuning gains in Table 2 (88.69% event classification, 70.72% temporal grounding) could reflect the model learning textual priors, dataset artifacts, or spurious correlations in the generated questions rather than genuine visual disaster monitoring. To support the central claim, the authors should (i) conduct a human evaluation of a random sample of image-caption pairs to measure what fraction of the described phenomena are actually visible in the corresponding Sentinel-2 frames; (ii) provide a text-only baseline (questions with no image input) to quantify how much of the reported accuracy is achievable without visual information; and (iii) report how cases with cloud cover, missing acquisitions, or events that fall between revisit dates are handled in label construction.
  2. [Section 4, Table 2] The temporal grounding task asks 'which date shows the first evidence of the [EVENT_TYPE]?' with the answer options being the dates of the satellite images in the sequence. The caption generation procedure produces descriptions for a list of event dates derived from the articles, not necessarily for Sentinel-2 acquisition dates. The paper does not specify how these caption dates are aligned to the actual image dates, nor how the ground-truth answer is selected when the article-reported event onset falls between two acquisitions. If the answer is simply the earliest image after the event onset date, the task may be partially solvable from the question text (which lists the dates) and the event type, without any visual evidence. Please describe the exact alignment algorithm, report the distribution of time gaps between the article-derived onset and the chosen image date, and break down the fine-tuned model's accuracy on subsets where the onset exactly matches an acquisition date versus where it does not. Such an analysis would clarify whether the model is truly detecting visual change or exploiting temporal priors.
  3. [Section 6.2, Table 4] The claim that fine-tuning 'improves the model's ability to connect language with visual features' is not well supported by the LLM-judge results. In Table 4, the fine-tuned model's 'Visual Evidence' score is 2.50, only modestly above TEOChat's 2.28, and its 'Overall' score (3.08) is identical to both VideoLLaVA and TEOChat. Since the open-ended responses are generated from the same news-derived captions used in training, the higher BLEU/ROUGE/METEOR scores in Table 3 could reflect better memorization of the training distribution rather than improved visual grounding. The authors should provide a finer-grained evaluation of the open-ended responses, for example per-question-type breakdowns or qualitative examples where the model cites a specific visual feature that is independently verified to be present in the image, to support the interpretation that the gains reflect visual understanding rather than textual prior matching.
minor comments (5)
  1. [Abstract] There is a punctuation issue: 'However, they remain limited by their narrow focus on specific disaster types, reliance on manual expert interpretation...' should be two sentences, e.g., '...analysis. However, they remain limited...'.
  2. [Section 3.2] The abstract states 'more than 10,000 FEMA disaster events' but Section 3.2 reports 9,996 incidents; please align the numbers.
  3. [Section 5] The fine-tuning details state that roughly 1/5 of the training set was used; please specify the exact number of QA pairs (approximately 8,862) and whether the subset was selected randomly or stratified by disaster type, since this affects the interpretation of the results.
  4. [Figure 3 and Section 3.1] The caption in Figure 3 says 'using the text alone' but the main text says 'This is done using the article content and dates alone.' Please clarify whether the dates provided to the LLM are the event dates from FEMA records or the Sentinel-2 acquisition dates, as this is important for understanding the temporal grounding label construction.
  5. [References] Reference [10] (Google Custom Search JSON API) is not cited in the main text where the Google Search API is mentioned; please add the citation at the point of use.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the fine-tuning evaluation is a standard train/test benchmark, and the acknowledged text-only label generation is a data-validity caveat rather than a circular derivation.

full rationale

The paper's central empirical claim is that fine-tuning TEOChat on MONITRS-QA improves performance on the benchmark's event classification, temporal grounding, and location grounding tasks. This is a conventional supervised setup: the model is trained on a training split and evaluated on a held-out event-disjoint test split, so the reported accuracies are not forced by construction. The training labels are not derived from the model's outputs, and the test labels are not derived from the training labels. Self-citations to the authors' prior work ([20], [21], [29]) appear only in the related-work discussion and are not load-bearing for the central result. No uniqueness theorem or ansatz is imported from prior work. The most concerning passage is in Section 3.1, where captions are generated 'using the article content and dates alone,' and the Limitations section concedes that 'the descriptions generated by LLMs based on news articles may not always accurately reflect what is visible in the satellite imagery.' This is a real validity threat: because the question text includes event types and dates, a model might partially exploit textual priors rather than visual evidence, and the benchmark cannot fully distinguish those strategies. However, this is an annotation-grounding and benchmark-validity concern, not a circular derivation. The evaluation does not reduce to the dataset-construction inputs by definition; it is an empirical question whether the model uses images, text, or both. Therefore the circularity score is low, with the caveat noted as a substantive limitation rather than as a circular step.

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

The paper contributes a dataset, so the main 'inputs' are annotation pipeline choices and external data sources rather than physical constants. The critical assumption is that text-derived captions are visually grounded; this is unverified and is the main source of correctness risk.

free parameters (4)
  • Satellite patch size = 5.12 x 5.12 km
    Chosen because it yields a 512x512 image at 10m resolution; this size determines which parts of an event are visible and is arbitrary.
  • Temporal buffer around event = 10 days before and after
    Added to ensure the full event is captured; the choice is ad hoc and affects which images are included.
  • Number of news articles per event = 5
    First five Google search results are used to bound API costs; coverage may be biased and incomplete.
  • Training subset fraction = 1/5 of training set
    Fine-tuning used only about 20% of the available training QA pairs due to computational constraints; results may improve with full data.
assumptions (4)
  • domain assumption FEMA disaster records accurately list the event type, county, and start/end dates.
    Used as ground truth for event classification and temporal grounding without external verification beyond the declarations.
  • domain assumption News articles returned by the search query are relevant and describe the event's location and progression factually.
    Articles are the sole source of locations and captions; no manual filtering beyond first-five selection.
  • ad hoc to paper LLM-generated captions from article text correspond to observable phenomena in Sentinel-2 imagery.
    Captions are written without viewing the images; this is the load-bearing premise for treating QA answers as visual ground truth.
  • domain assumption Sentinel-2 RGB imagery at 10m resolution and ~5 day revisit captures the disaster's visual evidence.
    Cloud cover, revisit gaps, and resolution may miss fast-evolving events; the paper acknowledges this in Limitations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MONITRS: Multimodal Observations of Natural Incidents Through Remote Sensing." pith.science (2026). https://pith.science/paper/WBXP5FNP

@misc{pith2026250716228,
  author       = {Pith},
  title        = {Pith review of: MONITRS: Multimodal Observations of Natural Incidents Through Remote Sensing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WBXP5FNP}},
  note         = {Machine review of arXiv:2507.16228}
}
read the original abstract

Natural disasters cause devastating damage to communities and infrastructure every year. Effective disaster response is hampered by the difficulty of accessing affected areas during and after events. Remote sensing has allowed us to monitor natural disasters in a remote way. More recently there have been advances in computer vision and deep learning that help automate satellite imagery analysis, However, they remain limited by their narrow focus on specific disaster types, reliance on manual expert interpretation, and lack of datasets with sufficient temporal granularity or natural language annotations for tracking disaster progression. We present MONITRS, a novel multimodal dataset of more than 10,000 FEMA disaster events with temporal satellite imagery and natural language annotations from news articles, accompanied by geotagged locations, and question-answer pairs. We demonstrate that fine-tuning existing MLLMs on our dataset yields significant performance improvements for disaster monitoring tasks, establishing a new benchmark for machine learning-assisted disaster response systems. Code can be found at: https://github.com/ShreelekhaR/MONITRS

Figures

Figures reproduced from arXiv: 2507.16228 by the authors.

Figure 1
Figure 1. Using news articles, we extract exact locations of disaster events and corresponding captions [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. We demonstrate the use of geocoded news articles used to capture a better understanding [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. We illustrate the captions generated through our dataset construction pipeline. After [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Our dataset represents the wide variety of natural disasters recorded by FEMA. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: B Prompts to LLM We use prompts to LLMs to act as language tools for two types of tasks in our work. The first being to read through and retrieve the relevant information from news articles to caption our image sequences, figures 6 and 7 The second being utilizing our …
Figure 5
Figure 5. Figure 5: Qualitative examples from both MONITRS and MONITRS-QA along with their respective [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Prompt given to LLM to extract proper nouns locations. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Prompt for creating chronological timelines of visually observable natural disaster events [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Prompt for generating multiple choice questions from natural disaster event statements [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Prompt for generating question-answer pairs from natural disaster event statements [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. A Multi-Agent Feedback System for Detecting and Describing News Events in Satellite Imagery

    cs.CV 2026-04 unverdicted novelty 7.0 of 10

    SkyScraper uses iterative multi-agent feedback to geocode news articles and synthesize captions for satellite image sequences, locating 5x more events than traditional methods and producing a new 5,000-sequence dataset.

  2. Changing Modalities: Adapting Remote Sensing Models to New Satellites and Sensors

    cs.CV 2026-06 unverdicted novelty 6.0 of 10

    DeluluNet enables continued prediction under modality substitution, addition, or subsets by training a multi-modal model from a unimodal teacher and unlabeled multimodal data via modality hallucination.

  3. A Multi-Agent Feedback System for Detecting and Describing News Events in Satellite Imagery

    cs.CV 2026-04 conditional novelty 6.0 of 10

    An iterative multi-agent feedback pipeline finds ~5× more multi-temporal news events in satellite imagery than traditional geocoding and yields a 5,000-sequence captioning dataset.

Reference graph

Works this paper leans on

49 extracted references · 42 canonical work pages · cited by 2 Pith papers

  1. [1]

    Meteor: An automatic metric for mt evaluation with improved correlation with human judgments

    Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65–72, 2005. 9

    work page 2005

  2. [2]

    Automated building damage assessment and large-scale mapping by integrating satellite imagery, gis, and deep learning

    Abdullah M Braik and Maria Koliou. Automated building damage assessment and large-scale mapping by integrating satellite imagery, gis, and deep learning. Computer-Aided Civil and Infrastructure Engineering, 39(15):2389–2404, 2024

    work page 2024

  3. [3]

    Application of uav remote sensing in natural disaster monitoring and early warning: an example of flood and mudslide and earthquake disasters

    Zhonghan Chen. Application of uav remote sensing in natural disaster monitoring and early warning: an example of flood and mudslide and earthquake disasters. Highlights in Science, Engineering and Technology, 85:924–933, 2024

    work page 2024

  4. [4]

    Changeclip: Remote sensing change detection with multimodal vision-language representation learning

    Sijun Dong, Libo Wang, Bo Du, and Xiaoliang Meng. Changeclip: Remote sensing change detection with multimodal vision-language representation learning. ISPRS Journal of Photogrammetry and Remote Sensing, 208:53–69, 2024

    work page 2024

  5. [5]

    Copernicus sentinel-2, 2024

    ESA. Copernicus sentinel-2, 2024

    work page 2024

  6. [6]

    National response framework, 2025

    FEMA. National response framework, 2025

    work page 2025

  7. [7]

    Openfema dataset, 2024

    Federal Emergency Management Agency (FEMA). Openfema dataset, 2024

    work page 2024

  8. [8]

    The use of high-resolution satellite topographic data to quantify volcanic activity at raung volcano (indonesia) from 2000 to 2021

    Federico Galetto, Diego Lobos Lillo, and Matthew E Pritchard. The use of high-resolution satellite topographic data to quantify volcanic activity at raung volcano (indonesia) from 2000 to 2021. Bulletin of Volcanology, 87(1):1–19, 2025

    work page 2000

Show all 49 references
  1. [10]

    Custom search json api

    Google Developers. Custom search json api. https://developers.google.com/custom-search/ v1/overview, 2025. Last updated: 2025-05-07, Accessed: 2025-05-16

    2025

  2. [11]

    xbd: A dataset for assessing building damage from satellite imagery

    Ritwik Gupta, Richard Hosfelt, Sandra Sajeev, Nirav Patel, Bryce Goodman, Jigar Doshi, Eric Heim, Howie Choset, and Matthew Gaston. xbd: A dataset for assessing building damage from satellite imagery. arXiv preprint arXiv:1911.09296, 2019

    1911 arXiv

  3. [12]

    Change captioning: A new paradigm for multitemporal remote sensing image analysis

    Genc Hoxha, Seloua Chouaf, Farid Melgani, and Youcef Smara. Change captioning: A new paradigm for multitemporal remote sensing image analysis. IEEE Transactions on Geoscience and Remote Sensing, 60:1–14, 2022

    2022

  4. [13]

    Teochat: A large vision-language assistant for temporal earth observation data

    Jeremy Andrew Irvin, Emily Ruoyu Liu, Joyce Chuyi Chen, Ines Dormoy, Jinyoung Kim, Samar Khanna, Zhuo Zheng, and Stefano Ermon. Teochat: A large vision-language assistant for temporal earth observation data. arXiv preprint arXiv:2410.06234, 2024

    arXiv

  5. [14]

    Geochat: Grounded large vision-language model for remote sensing

    Kartik Kuckreja, Muhammad Sohail Danish, Muzammal Naseer, Abhijit Das, Salman Khan, and Fa- had Shahbaz Khan. Geochat: Grounded large vision-language model for remote sensing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27831–27840, 2024

    2024

  6. [15]

    Vision-language models in remote sensing: Current progress and future trends

    Xiang Li, Congcong Wen, Yuan Hu, Zhenghang Yuan, and Xiao Xiang Zhu. Vision-language models in remote sensing: Current progress and future trends. IEEE Geoscience and Remote Sensing Magazine , 2024

    2024

  7. [16]

    Video-llava: Learning united visual representation by alignment before projection

    Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023

    2023 arXiv

  8. [17]

    Rouge: A package for automatic evaluation of summaries

    Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74–81, 2004

    2004

  9. [18]

    Remote sensing image change captioning with dual-branch transformers: A new method and a large scale dataset

    Chenyang Liu, Rui Zhao, Hao Chen, Zhengxia Zou, and Zhenwei Shi. Remote sensing image change captioning with dual-branch transformers: A new method and a large scale dataset. IEEE Transactions on Geoscience and Remote Sensing, 60:1–20, 2022

    2022

  10. [19]

    Remoteclip: A vision language foundation model for remote sensing

    Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  11. [20]

    Change-aware sampling and contrastive learning for satellite images

    Utkarsh Mall, Bharath Hariharan, and Kavita Bala. Change-aware sampling and contrastive learning for satellite images. In CVPR, 2023

    2023

  12. [21]

    Remote sensing vision-language foundation models without annotations via ground remote alignment

    Utkarsh Mall, Cheng Perng Phoo, Meilin Kelsey Liu, Carl V ondrick, Bharath Hariharan, and Kavita Bala. Remote sensing vision-language foundation models without annotations via ground remote alignment. ICLR, 2024. 10

    2024

  13. [22]

    Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data

    Oscar Manas, Alexandre Lacoste, Xavier Giró-i Nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9414–9423, 2021

    2021

  14. [23]

    Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data

    Oscar Manas, Alexandre Lacoste, Xavier Giró-i Nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In ICCV, 2021

    2021

  15. [24]

    Geocoding api

    Map Maker. Geocoding api. https://geocode.maps.co/, 2025. Accessed: 2025-05-16

    2025

  16. [25]

    Note on the sampling error of the difference between correlated proportions or percent- ages

    Quinn McNemar. Note on the sampling error of the difference between correlated proportions or percent- ages. Psychometrika, 12(2):153–157, 1947

    1947

  17. [26]

    Remote sensing change detection with transformers trained from scratch

    Mubashir Noman, Mustansar Fiaz, Hisham Cholakkal, Sanath Narayan, Rao Muhammad Anwer, Salman Khan, and Fahad Shahbaz Khan. Remote sensing change detection with transformers trained from scratch. IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  18. [27]

    Bleu: a method for automatic evaluation of machine translation

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002

    2002

  19. [28]

    National disaster management and monitoring using satellite remote sensing and geo-information

    Jongsoo Park, Hagyu Jeong, and Junwoo Lee. National disaster management and monitoring using satellite remote sensing and geo-information. Korean Journal of Remote Sensing, 40(5):813–832, 2024

    2024

  20. [29]

    Scale-aware recognition in satellite images under resource constraints

    Shreelekha Revankar, Cheng Perng Phoo, Utkarsh Mall, Bharath Hariharan, and Kavita Bala. Scale-aware recognition in satellite images under resource constraints. arXiv preprint arXiv:2411.00210, 2024

    2024 arXiv

  21. [30]

    The change you want to see

    Ragav Sachdeva and Andrew Zisserman. The change you want to see. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3993–4002, 2023

    2023

  22. [31]

    Natural disasters detection in social media and satellite imagery: a survey

    Naina Said, Kashif Ahmad, Michael Riegler, Konstantin Pogorelov, Laiq Hassan, Nasir Ahmad, and Nicola Conci. Natural disasters detection in social media and satellite imagery: a survey. Multimedia Tools and Applications, 78:31267–31302, 2019

    2019

  23. [32]

    Post-disaster recovery assessment with machine learning-derived land cover and land use information

    Mohammadreza Sheykhmousa, Norman Kerle, Monika Kuffer, and Saman Ghaffarian. Post-disaster recovery assessment with machine learning-derived land cover and land use information. Remote sensing, 11(10):1174, 2019

    2019

  24. [33]

    Flood detection in urban areas using satellite imagery and machine learning

    Ahad Hasan Tanim, Callum Blake McRae, Hassan Tavakol-Davani, and Erfan Goharian. Flood detection in urban areas using satellite imagery and machine learning. Water, 14(7):1140, 2022

    2022

  25. [34]

    Autonomous satellite wildfire detection using hyper- spectral imagery and neural networks: A case study on australian wildfire

    Kathiravan Thangavel, Dario Spiller, Roberto Sabatini, Stefania Amici, Sarathchandrakumar Thottuchirayil Sasidharan, Haytham Fayek, and Pier Marzocca. Autonomous satellite wildfire detection using hyper- spectral imagery and neural networks: A case study on australian wildfire...

    2023

  26. [35]

    Made to order: Discovering monotonic temporal changes via self-supervised video ordering

    Charig Yang, Weidi Xie, and Andrew Zisserman. Made to order: Discovering monotonic temporal changes via self-supervised video ordering. In European Conference on Computer Vision, pages 268–286. Springer, 2024

    2024

  27. [36]

    Video-llama: An instruction-tuned audio-visual language model for video understanding

    Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023

    2023 arXiv

  28. [37]

    Judging llm-as-a-judge with mt-bench and chatbot arena

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595–46623, 2023

    2023

  29. [38]

    Semantic-cc: Boosting remote sensing image change captioning via foundational knowledge and semantic guidance

    Yongshuo Zhu, Lu Li, Keyan Chen, Chenyang Liu, Fugen Zhou, and Zhenwei Shi. Semantic-cc: Boosting remote sensing image change captioning via foundational knowledge and semantic guidance. IEEE Transactions on Geoscience and Remote Sensing, 2024. 11 A Qualitative Results We incl...

    2024

  30. [39]

    Carefully review the attached articles about natural disasters and identify ONLY proper noun locations that are directly related to where the disaster occurred or had direct impact

  31. [40]

    Focus on extracting: • Specific sites where the event took place (cities, towns, neighborhoods) • Precise natural features affected (specific rivers, mountains, forests, beaches) • Particular infrastructure impacted (named dams, bridges, parks) • Exact regions directly experie...

  32. [41]

    Present your response in a simple string list format, with each location separated by a comma

  33. [42]

    If a location appears multiple times, include it only ONCE in your list

  34. [43]

    If the articles contain NO specific event locations, return only the word “no” (lower- case)

  35. [44]

    14 Task: Create a chronological timeline of observable natural disaster events from the provided news articles

    DO NOT include: • Broad geographical entities not directly affected (countries, states, unless the entire entity was impacted) • Locations only mentioned incidentally (headquarters of responding agencies, etc.) • Places mentioned for context but not directly experiencing the d...

  36. [45]

    Review the attached news articles for information about natural disasters (earthquakes, floods, hurricanes, wildfires, volcanic eruptions, etc.)

  37. [46]

    For each date in the provided list, identify natural disaster events that occurred on or by that date that would be seen remotely

  38. [47]

    Write a 1-2 sentence description for each date focusing specifically on the visible physical manifestations, such as: • Extent of flooding or inundation • Wildfire burn scars or active fire fronts • Hurricane cloud formations or aftermath flooding • Visible structural damage t...

  39. [48]

    If a specific date isn’t explicitly mentioned in the articles, use context clues to reason- ably infer when these visible changes occurred

  40. [49]

    Present your response as a simple chronological list with dates followed by descrip- tions

  41. [50]

    Emphasize the VISUAL aspects that would be detectable from above. Format example: June 15, 2023: Extensive flooding covered approximately 60 square miles of the Mississippi Delta region, with standing water clearly visible across previously inhabited areas and farmland. July 3...

    2023

Pith tools

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