Can We Predict the Unpredictable? Leveraging DisasterNet-LLM for Multimodal Disaster Classification

Arpita Soni; Gautam Siddharth Kashyap; Manaswi Kulahara; Nipun Joshi

REVIEW 5 major objections 6 minor 23 references

Can We Predict the Unpredictable? Leveraging DisasterNet-LLM for Multimodal Disaster Classification

T0 review · 5 major / 6 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 DisasterNet-LLM claims one fused model can classify disasters from news, images, and weather at 89.5% accuracy.

desk verdict Central accuracy figure is identical across two dissimilar dataset pairs and no data alignment is described; the paper is unverifiable as written. read the letter →

arxiv 2506.23462 v1 pith:J3NGEBHU submitted 2025-06-30 cs.LG cs.AI

Manaswi Kulahara , Gautam Siddharth Kashyap , Nipun Joshi , Arpita Soni This is my paper · ORCID

classification cs.LGcs.AI

keywords disasterclassificationlargelanguagemodelsmultimodalfusioncross-modalattentiongeospatialdataremotesensingmanagementadaptivetransformers

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

DisasterNet-LLM is an attempt to answer whether one large language model can classify disasters from all the data streams that arrive during an emergency: written reports and social posts, images, and geospatial or weather measurements. The paper claims that a single model fusing these modalities reaches 89.5% accuracy, an 88.0% F1 score, and an AUC of 0.92, beating the best classical baseline (SVM at 72.5% accuracy) by a wide margin. If the result holds, emergency managers could feed one model with satellite images, weather rasters, and news text and get more reliable, earlier disaster classification instead of maintaining separate pipelines. The paper also reports 91.65% accuracy on a Delhi urban-risk dataset covering fires, gas leaks, and building collapses, suggesting the approach transfers beyond benchmark imagery.

What carries the argument

The load-bearing object is the concatenated multimodal embedding $X = [T; I; G]$, where $T$ comes from a pretrained GPT text encoder, $I$ from CLIP's image encoder, and $G$ from a geospatial neural network (NN-GLS). Cross-modal attention then lets a token from one modality query keys and values drawn from all modalities, and a sigmoid-gated adaptive transformer computes weights $W_{\text{adapt}} = \sigma(W_a X + b_a)$ from the incoming data to emphasize the most relevant modality. The classification head is a softmax layer with cross-entropy loss. What does the work is the claim that this gated cross-modal fusion, rather than any single encoder, produces the high classification accuracy.

What would settle it

Run the described pipeline with the four public datasets using verified event-aligned multimodal samples and the stated hyperparameters; if the samples cannot be paired or accuracy drops toward single-modality levels once pairing is enforced, the 89.5% figure is not a valid multimodal result.

Watch

Extended reading notes

Core claim

The central claim is that a Transformer-based architecture can fuse textual, visual, and geospatial inputs into one embedding and use cross-modal attention to classify disasters better than any single-modality or classical model. In the paper's formulation, text embeddings from a pretrained GPT model, image embeddings from CLIP, and geospatial embeddings from a neural network are concatenated as $X = [T; I; G]$, then processed by cross-modal attention in which queries from one modality attend to keys and values from all modalities, and by an adaptive transformer whose gating weights are computed from the data itself. The final softmax layer produces the disaster class, trained with categorical cross-entropy. The paper reports that this pipeline achieves 89.50% accuracy, 88.00% F1, 0.92 AUC, 89.00% precision, 87.50% recall, and 0.88 BERTScore on the combined benchmarks, and 91.65% accuracy on an urban-risk dataset.

Load-bearing premise

The result stands on the assumption that the four datasets can be aligned into a single multimodal training set with shared labels, and that concatenating GPT, CLIP, and geospatial embeddings yields a compatible feature space.

Editorial extensions

If this is right

If the reported accuracy is reproducible, DisasterNet-LLM would beat the best classical benchmark in the paper by roughly 17 percentage points on the main image dataset.
The same architecture could be used in early-warning systems that ingest weather rasters and news text alongside imagery as an event unfolds.
The low error metrics (MAE 0.12 and RMSE 0.18) imply the fused representation may also support regression-style predictions such as damage severity, not only class labels.
The Delhi urban-risk results suggest fine-tuning on local incident registries lets the model handle domain-specific categories like gas leaks and building collapses.
A BERTScore of 0.88 suggests the text modality's semantic content survives fusion and could support explainability by linking classifications to source reports.

Reading between the lines

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

The paper never states how the four datasets are paired into multimodal samples; constructing event-aligned triples (news, image, and weather raster for the same disaster) and measuring the accuracy change when each modality is removed would test whether the fusion is genuinely multimodal.
The identical headline metrics reported in the two comparison tables suggest one model evaluation is being reported against both dataset pairs; per-dataset and per-class breakdowns, which the paper says it omitted, would show where the gain actually comes from.
A natural probe of the adaptive gating is to feed deliberately mismatched modalities, such as flood text with a fire image, and see whether the model's dynamic weights reweight the modalities sensibly or simply trust text.

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.

Desk Editor's Note

read the letter

If you come across this paper, treat the headline accuracy as unverified. The 89.50% accuracy and 88.00% F1 appear verbatim in both Table I (datasets [20] and [21]) and Table II (datasets [22] and [23]), which differ in size, class structure, and modality composition. That is either a transcription slip or the same result reused for two distinct evaluations—neither is acceptable evidence. On top of that, the paper never explains how the four datasets are aligned into the multimodal input X = [T; I; G] in Eq. (1). The news articles, ERA5 rasters, and images have no obvious pairing key. Without that, the evaluation is undefined.

What the paper does have: it concatenates embeddings from GPT, CLIP, and a geospatial network, runs them through a standard scaled-dot-product attention, and adds a sigmoid gate. That is a routine stack, not a new mechanism. The authors say they reran baselines to fill missing metrics, which is fair, but they give no protocol, code, or data, so the baseline numbers are as unverifiable as their own. There are also small but telling signs of sloppiness: the abstract reports 'AUC of 0.92%' and 'BERTScore of 0.88%'—the percentages are meaningless for these metrics—and MAE/RMSE are listed for a classification task without any definition of the target these errors are measured against. The architecture itself is a plain concatenation-plus-attention; calling it an 'adaptive transformer' does not make it one.

Who would get value from this? A practitioner looking for a deployable system will not find enough detail to reproduce anything. A researcher looking for architectural novelty will find nothing. The paper's only real use is as a cautionary example of opaque evaluation. My recommendation: desk reject. The central claim is unverifiable and internally suspicious. If the authors release code, aligned data, and a proper experimental protocol, the work might be worth a second look—but as it stands, it does not merit referee time.

Referee Report

5 major / 6 minor

Summary. The paper proposes DisasterNet-LLM, a multimodal deep-learning framework that concatenates text embeddings (GPT-style), image embeddings (CLIP), and geospatial embeddings (NN-GLS) and processes them through cross-modal attention and an adaptive transformer for disaster classification. The experiments compare the proposed model against classical baselines on image datasets [20] and [21], and against LLM/retrieval baselines on ERA5 and news datasets [22] and [23], reporting accuracy 89.50%, F1 88.00%, AUC 0.92, precision 89.00%, recall 87.50%, BERTScore 0.88, MAE 0.12, and RMSE 0.18. An additional urban-risk experiment on a NIDM Delhi dataset is described in Section IV-D, with 91.65% overall accuracy. The central claim is that DisasterNet-LLM significantly outperforms state-of-the-art methods on multimodal disaster classification.

Significance. If the empirical claims were valid and reproducible, the paper would report a substantial advance in multimodal disaster classification. The architecture is specified in closed form in Equations (1)-(7), which makes the modeling assumptions reviewable, and the datasets used are largely public. However, the experimental section does not provide a reproducible protocol: no data-alignment procedure, label definitions, baseline rerun code, confidence intervals, or statistical tests are given, and no code or aligned data is released. The apparent reuse of identical result rows across the two main comparison tables is a serious integrity concern. The work therefore does not, in its current form, substantiate its state-of-the-art claim.

major comments (5)

[IV-C, Tables I and II] The DisasterNet-LLM row in Table I and the DisasterNet-LLM row in Table II are numerically identical across all eight reported metrics (accuracy 89.50, F1 88.00, AUC 0.92, precision 89.00, recall 87.50, BERTScore 0.88, MAE 0.12, RMSE 0.18), even though the two tables claim to evaluate on different dataset pairs with different sizes, class structures, and modalities ([20]/[21] versus [22]/[23]). This is strong internal evidence that the reported performance was not obtained independently on the second pair, and it invalidates the multimodal comparison in Table II as presented.
[IV-A and Equation (1)] The paper never describes how the four datasets are aligned into a single multimodal training set with shared labels. Equation (1) defines X = [T; I; G] as a concatenation of text, image, and geospatial matrices, but the manuscript does not specify a pairing key that would allow a given news article, ERA5 raster, and image to form one joint sample. Without such a pairing procedure, the fused input X is not well defined, and the reported accuracy cannot be interpreted as multimodal classification performance.
[IV-B and IV-C, Table II] The comparison against Top-1 Sim, LLaVA Tuning, and CLLMate is not verifiable. The note in Section IV-B says 'we rerun those baseline models in order to compute those missing measures', but no rerun protocol, split, label set, or hyperparameter choices are reported. Since these baselines come from a weather/climate event forecasting paper [23], the manuscript must demonstrate that they were adapted to the same task and data; otherwise the comparison in Table II is not meaningful.
[IV-B and the abstract] The metric reporting is internally inconsistent: the abstract states 'AUC of 0.92%, and BERTScore of 0.88%', while the tables and Section IV-C report AUC = 0.92 and BERTScore = 0.88. AUC and BERTScore are not naturally expressed in percent units, and the percentage notation in the abstract is dimensionally wrong. In addition, no confidence intervals, standard deviations, or significance tests are provided for any metric, so the claim that the proposed model is 'significantly outperforming' existing methods is not statistically supported.
[IV-D and Section V] The NIDM urban-risk experiment is too underspecified to support any claim: no dataset size, label taxonomy, collection procedure, or train/validation/test split is given, and the map in Fig. 2 is not described. The concluding sentence admits that category-wise classification results were omitted 'due to page constraints', but the aggregate numbers alone do not permit verification of the 91.65% accuracy claim.

minor comments (6)

[Abstract] The phrases 'AUC of 0.92%' and 'BERTScore of 0.88%' should be corrected to 'AUC of 0.92' and 'BERTScore of 0.88', or the corresponding decimal units should be defined.
[III (Modeling)] Equation (2) invokes the standard Transformer attention formula but gives no citation for the Transformer architecture; a reference to the original Transformer paper should be added.
[III (Modeling)] The text refers to 'GPT 0' in the architecture figure caption; the implementation is not specified (which GPT model, whether it is frozen or fine-tuned), and the section should state the exact model names and parameter counts.
[IV-B] The note about missing baseline metrics is vague; the paper should identify which baselines lacked which metrics and describe the exact rerun protocol, including data splits and hyperparameters.
[IV-A] The ERA5 dataset is cited to [22], but the description of a four-variable raster dataset spanning July 2015 to June 2023 is not directly the content of that reference; the authors should clarify the exact ERA5 product, variables, and preprocessing steps.
[IV-D, Fig. 2] Figure 2 is labeled only 'Base map of Delhi' and is not discussed in the text; the figure should either be referenced in detail or removed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the performance claim is an empirical result, not derived from the model's own definitions.

full rationale

DisasterNet-LLM is an empirical evaluation paper; the claimed 89.50% accuracy and 88.00% F1 score are measured results on held-out test splits, not logical consequences of the model definitions. Equations (1)-(7) define a standard multimodal classifier through concatenation, cross-modal attention, adaptive gating, softmax, and categorical cross-entropy; none of these equations defines the reported accuracy in terms of itself, and no parameter is fitted to the test metric and then renamed as a prediction. The paper contains no self-citations by the present authors and invokes no uniqueness theorem. The closest concern, the numerically identical DisasterNet-LLM rows in Tables I and II across different dataset pairs, is an evaluation-validity and reproducibility issue (possibly indicating a reused result), not a circularity: it does not show that the reported accuracy reduces by construction to the training labels or to the method's own assumptions. The omitted category-wise results and the unspecified multimodal alignment procedure are completeness and verification limitations, not circular reasoning. The empirical comparison is self-contained relative to external datasets and baselines, so the circularity score is 0.

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

The paper introduces no hand-fitted constants or new physical entities. It relies on off-the-shelf pretrained models (GPT, CLIP, NN-GLS) and assumes their embeddings can be concatenated for disaster classification. The central assumptions are that the datasets are properly aligned and labeled and that standard attention transfers across modalities.

assumptions (4)

domain assumption Pretrained GPT, CLIP, and NN-GLS embeddings faithfully represent disaster-related text, images, and geospatial data.
The paper builds on off-the-shelf models without domain adaptation or fine-tuning for disaster semantics (Sec III-A).
domain assumption Text, image, and geospatial samples can be concatenated into a single vector without loss of modality-specific structure.
Equation (1) defines X = [T; I; G] with no alignment or weighting, assuming a shared embedding space.
standard math Standard scaled dot-product attention is applicable across heterogeneous modalities.
Equation (2) reuses the transformer attention formula without justification for cross-modal inputs.
domain assumption The datasets [20]-[23] provide correctly labeled samples for the same disaster events.
No labeling or alignment protocol is described in Sec IV-A.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Can We Predict the Unpredictable? Leveraging DisasterNet-LLM for Multimodal Disaster Classification." pith.science (2026). https://pith.science/paper/J3NGEBHU

@misc{pith2026250623462,
  author       = {Pith},
  title        = {Pith review of: Can We Predict the Unpredictable? Leveraging DisasterNet-LLM for Multimodal Disaster Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J3NGEBHU}},
  note         = {Machine review of arXiv:2506.23462}
}

read the original abstract

Effective disaster management requires timely and accurate insights, yet traditional methods struggle to integrate multimodal data such as images, weather records, and textual reports. To address this, we propose DisasterNet-LLM, a specialized Large Language Model (LLM) designed for comprehensive disaster analysis. By leveraging advanced pretraining, cross-modal attention mechanisms, and adaptive transformers, DisasterNet-LLM excels in disaster classification. Experimental results demonstrate its superiority over state-of-the-art models, achieving higher accuracy of 89.5%, an F1 score of 88.0%, AUC of 0.92%, and BERTScore of 0.88% in multimodal disaster classification tasks.

Figures

Figures reproduced from arXiv: 2506.23462 by the authors.

**Figure 1.** shows the architecture consists of a hybrid framework combining LLMs (GPT0 ) for textual data, CLIP1 for image data, and Geospatial Neural Networks (NN-GLS) [19] for spatial data. Below, we present the mathematical formulation of each of the core components of DisasterNet-LLM. A. Multimodal Data Fusion Module Let T = {t1, t2, ..., tN } denote the set of textual data (social media posts, reports), I = {i1, i2, ..., i… view at source ↗

**Figure 2.** Base map of Delhi V. CONCLUSION AND FUTURE WORKS In this study, we proposed DisasterNet-LLM, a robust multimodal framework for disaster prediction. The model demonstrated superior performance across multiple metrics, outperforming existing SOTA approaches. Its ability to integrate textual and visual modalities effectively enables comprehensive decision-making in disaster management. For future work, we aim to enha… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 22 canonical work pages

[20]

A novel disaster image data-set and characteristics analysis using attention model,

F. F. Niloy, A. B. S. Nayem, A. Sarker, O. Paul, M. A. Amin, A. A. Ali, M. I. Zaber, A. M. Rahman et al. , “A novel disaster image data-set and characteristics analysis using attention model,” in 2020 25th International Conference on Pattern Recognition (ICPR). IEEE, 2021, pp. 6116–6122

work page 2020
[21]

Medic: a multi-task learning dataset for disaster image classification,

F. Alam, T. Alam, M. A. Hasan, A. Hasnat, M. Imran, and F. Ofli, “Medic: a multi-task learning dataset for disaster image classification,” Neural Computing and Applications , vol. 35, no. 3, pp. 2609–2632, 2023

work page 2023
[22]

The era5 global reanalysis,

H. Hersbach, B. Bell, P. Berrisford, S. Hirahara, A. Hor ´anyi, J. Mu ˜noz-Sabater, J. Nicolas, C. Peubey, R. Radu, D. Schepers et al. , “The era5 global reanalysis,” Quarterly Journal of the Royal Meteorological Society, vol. 146, no. 730, pp. 1999–2049, 2020

work page 1999
[23]

Cllmate: A multimodal llm for weather and climate events forecasting,

H. Li, Z. Wang, J. Wang, A. K. H. Lau, and H. Qu, “Cllmate: A multimodal llm for weather and climate events forecasting,” arXiv preprint arXiv:2409.19058 , 2024

arXiv 2024
[1]

Toward scalable damage assessment for rapid disaster re- sponse,

M. Wieland, V . Hertel, C. Geiss, S. Martinis, and K. Lechner, “Toward scalable damage assessment for rapid disaster re- sponse,” in IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium . IEEE, 2024, pp. 3902–3905

work page 2024
[2]

Uncertainty- aware 2d/3d change detection for natural disaster response,

C. H ¨ummer, D. Lallement, and D. Youssefi, “Uncertainty- aware 2d/3d change detection for natural disaster response,” in IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2024, pp. 3796–3801

work page 2024
[3]

Social media for emergency rescue: An analysis of rescue requests on twitter during hurricane har- vey,

L. Zou, D. Liao, N. S. Lam, M. A. Meyer, N. G. Gharaibeh, H. Cai, B. Zhou, and D. Li, “Social media for emergency rescue: An analysis of rescue requests on twitter during hurricane har- vey,” International Journal of Disaster Risk Reduction , vol. 85, p. 103513, 2023

work page 2023
[4]

Detecting natural hazard-related disaster impacts with social media analyt- ics: the case of australian states and territories,

T. Yigitcanlar, M. Regona, N. Kankanamge, R. Mehmood, J. D’Costa, S. Lindsay, S. Nelson, and A. Brhane, “Detecting natural hazard-related disaster impacts with social media analyt- ics: the case of australian states and territories,” Sustainability, vol. 14, no. 2, p. 810, 2022

work page 2022

Show all 23 references

[5]

Venice was flooding... one tweet at a time,

V . Lorini, P. Rufolo, and C. Castillo, “Venice was flooding... one tweet at a time,” Proceedings of the ACM on Human-Computer Interaction, vol. 6, no. CSCW2, pp. 1–16, 2022

2022
[6]

Damage assessment from social media imagery data during disasters,

D. T. Nguyen, F. Ofli, M. Imran, and P. Mitra, “Damage assessment from social media imagery data during disasters,” in Proceedings of the 2017 IEEE/ACM international conference on advances in social networks analysis and mining 2017 , 2017, pp. 569–576

2017
[7]

Enhancing disaster response with automated text information extraction from social media images,

H. B. Firmansyah, J. L. Fernandez-Marquez, J. Cerquides, V . Lorini, C. A. Bono, and B. Pernici, “Enhancing disaster response with automated text information extraction from social media images,” in 2023 IEEE Ninth International Conference on Big Data Computing Service and App...

2023
[8]

Crowd4ems: A crowdsourcing platform for gathering and geolocating social media content in disaster response,

A. Ravi Shankar, J. L. Fernandez-Marquez, B. Pernici, G. Scalia, M. R. Mondardini, and G. Di Marzo Serugendo, “Crowd4ems: A crowdsourcing platform for gathering and geolocating social media content in disaster response,” Inter- national Archives of the Photogrammetry, Remote S...

2019
[9]

Disaster early warning and damage as- sessment analysis using social media data and geo-location information,

D. Wu and Y . Cui, “Disaster early warning and damage as- sessment analysis using social media data and geo-location information,” Decision support systems , vol. 111, pp. 48–59, 2018

2018
[10]

Domain knowledge- aware remote sensing foundation model for flood detection in multi-spectral imagery,

Y . Li, B. Dang, F. Wei, J. Tan, and Y . Lin, “Domain knowledge- aware remote sensing foundation model for flood detection in multi-spectral imagery,” in IGARSS 2024-2024 IEEE Interna- tional Geoscience and Remote Sensing Symposium . IEEE, 2024, pp. 785–789

2024
[11]

Spacenet 8: Winning approaches to multi-class feature segmentation from satellite imagery for flood disasters,

R. H ¨ansch, J. Arndt, D. Lunga, T. Pedelose, A. Boedihardjo, J. Pfefferkorn, D. Petrie, and T. M. Bacastow, “Spacenet 8: Winning approaches to multi-class feature segmentation from satellite imagery for flood disasters,” in IGARSS 2023-2023 IEEE International Geoscience and R...

2023
[12]

Large language model applications for evaluation: Opportunities and ethical implications,

C. B. Head, P. Jasper, M. McConnachie, L. Raftree, and G. Higdon, “Large language model applications for evaluation: Opportunities and ethical implications,” New directions for evaluation, vol. 2023, no. 178-179, pp. 33–46, 2023

2023
[13]

Parameter-efficient fine-tuning of large-scale pre-trained language models,

N. Ding, Y . Qin, G. Yang, F. Wei, Z. Yang, Y . Su, S. Hu, Y . Chen, C.-M. Chan, W. Chen et al. , “Parameter-efficient fine-tuning of large-scale pre-trained language models,” Nature Machine Intelligence, vol. 5, no. 3, pp. 220–235, 2023

2023
[14]

Knowledge injection to counter large language model (llm) hallucination,

A. Martino, M. Iannelli, and C. Truong, “Knowledge injection to counter large language model (llm) hallucination,” in Euro- pean Semantic Web Conference. Springer, 2023, pp. 182–185

2023
[15]

Health system-scale language models are all-purpose predic- tion engines,

L. Y . Jiang, X. C. Liu, N. P. Nejatian, M. Nasir-Moin, D. Wang, A. Abidin, K. Eaton, H. A. Riina, I. Laufer, P. Punjabi et al. , “Health system-scale language models are all-purpose predic- tion engines,” Nature, vol. 619, no. 7969, pp. 357–362, 2023

2023
[16]

Fine-tuning gpt-3 for legal rule classification,

D. Liga and L. Robaldo, “Fine-tuning gpt-3 for legal rule classification,” Computer Law & Security Review , vol. 51, p. 105864, 2023

2023
[17]

Dsqa-llm: domain- specific intelligent question answering based on large language model,

D. Huang, Z. Wei, A. Yue, X. Zhao, Z. Chen, R. Li, K. Jiang, B. Chang, Q. Zhang, S. Zhang et al. , “Dsqa-llm: domain- specific intelligent question answering based on large language model,” in International Conference on AI-generated Content . Springer, 2023, pp. 170–180

2023
[18]

Enhancing emergency decision-making with knowledge graphs and large language models,

M. Chen, Z. Tao, W. Tang, T. Qin, R. Yang, and C. Zhu, “Enhancing emergency decision-making with knowledge graphs and large language models,” International Journal of Disaster Risk Reduction, vol. 113, p. 104804, 2024

2024
[19]

Neural networks for geospatial data,

W. Zhan and A. Datta, “Neural networks for geospatial data,” Journal of the American Statistical Association , no. just- accepted, pp. 1–21, 2024

2024

Pith tools

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

[1] [20]

A novel disaster image data-set and characteristics analysis using attention model,

F. F. Niloy, A. B. S. Nayem, A. Sarker, O. Paul, M. A. Amin, A. A. Ali, M. I. Zaber, A. M. Rahman et al. , “A novel disaster image data-set and characteristics analysis using attention model,” in 2020 25th International Conference on Pattern Recognition (ICPR). IEEE, 2021, pp. 6116–6122

work page 2020

[2] [21]

Medic: a multi-task learning dataset for disaster image classification,

F. Alam, T. Alam, M. A. Hasan, A. Hasnat, M. Imran, and F. Ofli, “Medic: a multi-task learning dataset for disaster image classification,” Neural Computing and Applications , vol. 35, no. 3, pp. 2609–2632, 2023

work page 2023

[3] [22]

The era5 global reanalysis,

H. Hersbach, B. Bell, P. Berrisford, S. Hirahara, A. Hor ´anyi, J. Mu ˜noz-Sabater, J. Nicolas, C. Peubey, R. Radu, D. Schepers et al. , “The era5 global reanalysis,” Quarterly Journal of the Royal Meteorological Society, vol. 146, no. 730, pp. 1999–2049, 2020

work page 1999

[4] [23]

Cllmate: A multimodal llm for weather and climate events forecasting,

H. Li, Z. Wang, J. Wang, A. K. H. Lau, and H. Qu, “Cllmate: A multimodal llm for weather and climate events forecasting,” arXiv preprint arXiv:2409.19058 , 2024

arXiv 2024

[5] [1]

Toward scalable damage assessment for rapid disaster re- sponse,

M. Wieland, V . Hertel, C. Geiss, S. Martinis, and K. Lechner, “Toward scalable damage assessment for rapid disaster re- sponse,” in IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium . IEEE, 2024, pp. 3902–3905

work page 2024

[6] [2]

Uncertainty- aware 2d/3d change detection for natural disaster response,

C. H ¨ummer, D. Lallement, and D. Youssefi, “Uncertainty- aware 2d/3d change detection for natural disaster response,” in IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2024, pp. 3796–3801

work page 2024

[7] [3]

Social media for emergency rescue: An analysis of rescue requests on twitter during hurricane har- vey,

L. Zou, D. Liao, N. S. Lam, M. A. Meyer, N. G. Gharaibeh, H. Cai, B. Zhou, and D. Li, “Social media for emergency rescue: An analysis of rescue requests on twitter during hurricane har- vey,” International Journal of Disaster Risk Reduction , vol. 85, p. 103513, 2023

work page 2023

[8] [4]

Detecting natural hazard-related disaster impacts with social media analyt- ics: the case of australian states and territories,

T. Yigitcanlar, M. Regona, N. Kankanamge, R. Mehmood, J. D’Costa, S. Lindsay, S. Nelson, and A. Brhane, “Detecting natural hazard-related disaster impacts with social media analyt- ics: the case of australian states and territories,” Sustainability, vol. 14, no. 2, p. 810, 2022

work page 2022