PaperlayerView exact revision on arXiv

Source-linked AI summary

Harnessing Large Language Models for Disaster Management: A Survey

Zhenyu Lei, Yushun Dong, Weiyu Li, Rong Ding, Qi Wang, Jundong Li

arXiv:2501.06932v2cs.CLcs.CYcs.LG

TL;DR

Research on LLMs for natural disaster management lacks a systematic review, leaving the field’s applications and open challenges without a consolidated survey. This paper reviews applications across disaster phases, organizes them in a taxonomy, and compiles datasets and challenges; it reports that current work is concentrated on response and often uses generic LLMs. The survey’s scope is limited to existing LLM applications, mostly for natural rather than man-made disasters.

  • Problem

    A systematic review of LLMs in disaster management is absent, limiting researchers’ and practitioners’ ability to identify best practices, research gaps, and ways to optimize deployment.

    Source: §1 · p. 1

  • Method

    The paper surveys LLM applications across four disaster phases, classifies them by scenarios, tasks, and model architectures, and compiles datasets and challenges.

    Source: §1 · p. 1, §3 · p. 2

  • Results

    The survey reports that current applications are concentrated on disaster response and often use generic LLMs, while other phases remain underexplored.

    Source: §5 · p. 8, §A.3 · p. 18

  • Takeaways & Limitations

    The survey aims to guide collaborative work by AI researchers and decision-makers toward LLM-driven disaster management and more resilient communities.

    Source: §6 · p. 9

    The survey covers existing LLM applications, focuses mostly on natural rather than man-made disasters, and excludes unexplored scenarios such as recovery repair-cost evaluation.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Large language models (LLMs) have revolutionized scientific research with their exceptional capabilities and transformed various fields. Among their practical applications, LLMs have been playing a crucial role in mitigating threats to human life, infrastructure, and the environment. Despite growing research in disaster LLMs, there remains a lack of systematic review and in-depth analysis of LLMs for natural disaster management. To address the gap, this paper presents a comprehensive survey of existing LLMs in natural disaster management, along with a taxonomy that categorizes existing works based on disaster phases and application scenarios. By collecting public datasets and identifying key challenges and opportunities, this study aims to guide the professional community in developing advanced LLMs for disaster management to enhance the resilience against natural disasters.

1 Introduction

Natural disasters threaten people, infrastructure, and the environment, while LLM capabilities may support disaster-management tasks. The paper addresses the lack of a systematic review with a four-phase survey, taxonomy, resource compilation, and discussion of challenges and future directions.

1 Introduction

  • Natural disasters pose threats to human life, infrastructure, and the environment, motivating effective disaster-management strategies.The introduction describes LLM capabilities in contextual understanding, reasoning, and complex problem-solving across modalities.Source: §1 · p. 1
  • LLMs are described as supporting analysis of real-time disaster data, communication with affected communities, and critical decision-making.These are presented as potential disaster-management roles.Source: §1 · p. 1
  • The paper addresses the absence of a systematic review of LLMs in disaster management by surveying their applications across four disaster phases.The authors identify this gap as limiting researchers and practitioners in finding best practices, research gaps, and ways to optimize LLM deployment.Source: §1 · p. 1
  • The survey’s taxonomy integrates application scenarios, specific tasks, and model architectures for disaster-related challenges.It covers encoder-based, (encoder-)decoder, and multimodal LLMs across mitigation, preparedness, response, and recovery.Source: §1 · p. 1
  • The paper compiles public resources, including datasets, and identifies challenges and future research directions for LLM-driven disaster management.It aims to guide AI researchers, policymakers, and practitioners.Source: §1 · p. 1

2 Background

The background presents disaster management as a multidisciplinary effort to reduce immediate harm and support long-term resilience and recovery. It describes four interconnected phases and outlines three LLM architecture types, common tasks, and adaptation techniques.

2 Background

  • Disaster management comprises four interconnected phases: mitigation, preparedness, response, and recovery.They cover risk reduction, readiness, immediate needs, and rebuilding after disasters, respectively.Source: §2 · p. 2
  • Disaster management integrates resources, expertise, and strategies to minimize immediate damage while fostering long-term resilience and adaptive recovery.Source: §2 · p. 2
  • The paper groups LLMs into encoder-based, (encoder-)decoder, and multimodal architectures, and lists classification, estimation, extraction, and generation as common downstream tasks.Examples include BERT, GPT, damage classification, severity estimation, knowledge extraction, and report generation.Source: §2 · p. 2
  • Fine-tuning and prompting are identified as common techniques for tailoring LLMs to disaster-management tasks.Source: §2 · p. 2

3 LLM For Disaster Management

This section surveys LLM applications across disaster-management phases and scenarios, covering their classification by task and model architecture as well as uses in recovery planning, progress tracking, and public-sentiment assessment.

3.1 Disaster Mitigation

  • The taxonomy spans four disaster phases; in mitigation, LLMs assess infrastructure and community vulnerabilities, classify infrastructure-failure concerns, and answer community queries using the Social Vulnerability Index.The Infrastructure Ombudsman combines supervised encoder-based learning and zero-shot prompting to classify concerns from social media; (encoder-)decoder LLMs retrieve and use the SVI for vulnerability queries (Martelo and Wang, 2024).Source: §3 · p. 2, §3.1 · p. 2

3.2 Disaster Preparedness

  • Preparedness work extracts disaster knowledge for public awareness, uses it for question answering, and applies LLMs to prediction, warnings, and evacuation planning.NER and instructional fine-tuning support knowledge extraction and graph construction, while RAG can improve knowledge integration. Prediction draws on text, geographic descriptions, satellite imagery, and retrieved historical flood data (Zeng and Bertsimas, 2023; Liu and Zhong, 2023).Source: §3.2 · p. 2, §3.2.1 · pp. 2–3, §3.2.2 · p. 3

3.2.3 Disaster Warning

3.3.1 Disaster Identification

  • Disaster identification uses fine-tuned encoder LLMs, ensembles, hybrid neural architectures, active learning, prompts, and multimodal methods to classify disaster-related content in social posts and other data (Ningsih and Hadiana, 2021; Singh et al., 2022; Lamsal et al., 2024a; Mukhtiar et al., 2023; Paul et al., 2023; Taghian Dinani et al., 2023).Hybrid methods add CNNs, attention-based BiLSTMs, or GNNs; multimodal approaches combine text and images using aggregation, attention, Cycle-GAN fusion, domain adaptation, or ensembles (Franceschini et al., 2024; Huang et al., 2022; Manthena, 2023; Shetty et al., 2024; Zhou et al., 2023b; Wang and Wang, 2022; Hanif et al., 2023).Source: §3.3.1 · p. 4, Relevance Classification with… · p. 4

3.3.2 Disaster Situation Assessment

  • Situation assessment ranges from classifying situational posts and locations to estimating earthquake intensity or flood depth and generating textual reports from images (Madichetty and Sridevi, 2021; Raj et al., 2023; Kanth et al., 2022; Mousavi et al., 2024; Hu and Rahnemoonfar, 2024; Wolf et al., 2023).CoT reasoning supports earthquake intensity estimates in Modified Mercalli Intensity, while FloodDepth-GPT estimates floodwater depth from images with prompt-based guidance.Source: §3.3.2 · pp. 4–5

3.3.3 Disaster Information Coordination

  • Information coordination filters and classifies posts, identifies needs and locations, and supports summaries and reports, though usefulness-rating methods require a predefined relevance threshold (Toraman et al., 2023; Mehmood et al., 2024; Colverd et al., 2023).Encoder LLMs also match requests with resource offers using embedding-based retrieval; summarization methods include critical-post selection and multi-step query-focused generation, while report generation uses RAG or CoT.Source: §3.3.3 · pp. 5–6

3.3.4 Disaster Rescuing

  • For rescue and consultation, LLMs generate actionable plans, translate verbal inputs into robot commands, and provide guidance through document-grounded or multimodal question answering (Goecks and Waytowich, 2023; Panagopoulos et al., 2024; Xie et al., 2024; Chen et al., 2024; Xia et al., 2024; Liu et al., 2024).Consultation systems use RAG with verified disaster documents, knowledge-graph prompt chains, or fine-tuning alongside RAG; multimodal systems include VQA and tools combining vision-language, object detection, and segmentation models.Source: §3.3.4 · p. 6, §3.3.5 · p. 6

3.4 Disaster Recovery

  • Recovery applications assess physical damage and social sentiment, estimate severity, answer infrastructure-status questions, and use assessments to formulate recovery plans (Ziaullah et al., 2024).Methods include damage classification and fine-grained severity scores, RAG-based status retrieval, image-based assessment with GPT-4o or VQA, and few-shot extraction of fatality information.Source: §3.4.1 · p. 7, §3.4.2 · p. 7

3.4.2 Recovery Plan Generation

  • ChatGPT has been prompted to develop disaster recovery strategies for business restoration.Source: §3.4.2 · p. 7

3.4.3 Recovery Process Tracking

  • Continuous tracking checks recovery progress against the planned timeline so decision-makers can adapt strategies to evolving needs.Source: §3.4.3 · p. 7
  • Encoder-based LLMs such as BERTweet assess public sentiment during recovery, helping responders tailor efforts to affected populations’ emotional needs (CONTRERAS et al.).Source: §3.4.3 · p. 7

4 Datasets

The paper describes datasets used to evaluate disaster-management LLMs across classification, estimation, generation, and extraction. It notes their input and label types, evaluation metrics, and a publication-count figure comparing use of existing LLMs with development of new models across disaster phases.

4 Datasets

  • Figure 3 presents publication counts for work using existing LLMs versus developing new models across the four disaster-management phases.Source: Figure 3 · p. 8
  • The paper provides a comprehensive list of publicly available datasets in Appendix B.It states that disaster-related datasets have been used to evaluate LLMs in disaster management.Source: §4 · pp. 7–8
  • Classification datasets cover informativeness, humanitarian types, damage levels, and other attributes, using textual and sometimes visual inputs.Performance is typically evaluated with accuracy and F1 score.Source: §4 · p. 8
  • Estimation datasets provide quantitative labels such as flood depths and use metrics such as Mean Absolute Error.Source: §4 · p. 8
  • Generation datasets include question-answering pairs and document collections with expert-curated summaries; some question-answering datasets also include disaster images.Both question-answering and summarization tasks are evaluated using metrics such as BLEU.Source: §4 · p. 8
  • Extraction datasets label sentence elements such as keywords and locations, with tokens marked as outside, start, or end.They are primarily used for token-level classification and evaluated with classification metrics.Source: §4 · p. 8

5 Challenges and Opportunities

LLM disaster-management research faces domain-specific, phase-coverage, dataset, efficiency, generation-reliability, and evaluation challenges. The paper identifies approaches and opportunities to address these constraints.

Challenges and Opportunities

  • Generic LLMs are often treated as universal solutions, while disaster-specific challenges and tailored frameworks receive insufficient attention.Current applications also concentrate heavily on response, leaving other disaster phases underexplored.Source: §5 · p. 8
  • Disaster datasets skew toward classification, while uncertainty and bias in raw data complicate reliable dataset construction.Synthetic data generation is identified as a possible way to improve coverage across diverse disaster scenarios.Source: §5 · p. 8
  • Large-scale LLM efficiency challenges limit real-time emergency use, while lightweight models may compromise robustness in disaster-related tasks.The paper calls for balancing efficiency and reliability.Source: §5 · p. 8
  • (Encoder-)decoder LLM hallucinations can produce factually inaccurate outputs, including false evacuation routes and resource misallocation.RAG, domain-specific training, and uncertainty estimation are proposed as mitigation strategies.Source: §5 · p. 8
  • Expert reference sets provide high-quality supervision, but differing annotation criteria across datasets make published generation results difficult to compare directly.The paper calls for a unified evaluation protocol to support more reliable comparisons.Source: §5 · pp. 8–9

6 Conclusion

The paper surveys LLM applications across the four disaster phases and introduces a taxonomy of scenarios, tasks, and model architectures. It compiles public datasets and identifies challenges to inform collaborative work toward resilient communities and proactive disaster management.

Conclusion

  • The survey organizes LLM applications across four disaster phases using a taxonomy that connects application scenarios, specific tasks, and model architectures.It presents publicly available datasets and identifies key challenges.Source: §6 · p. 9
  • The authors aim to inspire collaboration between AI researchers and decision-makers on disaster-management applications.The stated goal is to support resilient communities and proactive disaster-management practices.Source: §6 · p. 9
  • The paper frames its survey and identified challenges as resources for developing LLM applications in disaster management.Its conclusion links this effort to realizing LLM potential for resilience and proactive practice.Source: §6 · p. 9

Limitations

The paper limits its survey to existing LLM applications, focuses mainly on natural disasters, and acknowledges constraints in categorization and dataset selection. Appendix materials summarize surveyed papers and illustrate model pipelines.

Limitations

  • The survey covers disaster-management applications where LLMs have already been used, focuses mostly on natural rather than man-made disasters, and omits unexplored scenarios such as recovery repair-cost evaluation.Its dataset compilation includes only a subset, prioritizing accessible resources; many datasets are closed, restricted, or of uncertain quality.Source: Limitations · p. 9

A.1 Summary Table

  • Table 1 summarizes surveyed papers by disaster phase, application scenario, specific task, and architecture type.Source: §A.1 · p. 18

A.2 Pipeline Illustration

  • Figure 4 illustrates encoder-based, (encoder-)decoder, and multimodal LLM pipelines across classification, extraction, estimation, and generation tasks.The figure presents their mechanisms and applications in disaster management.Source: §A.2 · p. 18

A.3 Statistics

The surveyed literature shows a gap between NLP and disaster-management communities. Publications are mostly application-focused, concentrate on response, and appear relatively infrequently in NLP- or AI-specific venues.

A.3 Statistics

  • The surveyed statistics indicate a significant gap between the NLP and disaster-management communities.The paper argues for stronger interdisciplinary collaboration to bridge these fields.Source: §A.3 · p. 18
  • Most surveyed studies are application-focused, primarily using existing LLMs rather than developing new frameworks.The majority rely on fine-tuning or prompting existing models for disaster-management tasks instead of designing novel architectures.Source: §A.3 · p. 18
  • Research remains concentrated on disaster response, with limited exploration of other critical disaster-management scenarios.Source: §A.3 · p. 18
  • Relatively few disaster-management papers appear in NLP- or AI-specific conferences and journals, reflecting limited engagement from the LLM research community.The paper calls for greater awareness and collaboration in the field.Source: §A.3 · p. 18

B Datasets

The appendix summarizes publicly available disaster-management datasets and excludes some single-disaster classification datasets already covered by comprehensive benchmarks.

B Datasets

  • Table 2 summarizes publicly available datasets used in disaster management.Source: B Datasets · p. 18
  • For classification, single-disaster datasets are excluded when they are already included in comprehensive benchmarks such as CrisisBench.Source: B Datasets · p. 18

B.1 Classification Datasets

Classification datasets cover relevance, humanitarian, and other crisis-message labels across events and formats. They range from event-specific collections to larger benchmarks, multilingual transfer sets, and multimodal data.

B.1 Classification Datasets

  • CrisisLexT6 covers six crisis events; CrisisLexT26 expands coverage to 26 events and six humanitarian categories, while CrisisNLP spans disaster-response and health-emergency classes.Source: §B.1 · pp. 18–19
  • SWDM2013 and ISCRAM2013 cover Joplin and Hurricane Sandy tweets, with ISCRAM2013 adding humanitarian labels; EPIC also supports relevance and humanitarian classification.Source: §B.1 · p. 19
  • CrisisMMD provides 16k labeled tweets and images from seven natural disasters, annotated for relevance, humanitarian categories, and damage severity.Source: §B.1 · p. 19
  • Multi-Crisis combines seven datasets spanning 53 crisis events and nine domains for transfer across high- and low-resource languages and unseen crisis domains.Source: §B.1 · p. 19
  • CrisisBench consolidates nine datasets into a benchmark for relevance and humanitarian classification.Source: §B.1 · p. 19
  • Eyewitness Messages classifies 14,000 tweets from earthquakes, hurricanes, and wildfires as direct, indirect, or vulnerable eyewitness accounts.Source: §B.1 · p. 19
  • TREC Incident Streams labels 20k tweets for actionable information and criticality, while DSM contains 10K relevance-labeled tweets.Source: §B.1 · p. 19
  • HumAID contains 77k labeled tweets sampled from 24 million tweets across 19 disasters, balanced by disaster type and labeled in seven humanitarian categories.Source: §B.1 · p. 19

B.2 Estimation Datasets

Estimation datasets cover earthquake intensity, flood depth, and usefulness ratings; extraction datasets support location, event-information, entity-relation, and rationale tasks.

B.2 Estimation Datasets

  • Did You Feel It records ground-shaking intensity and geographic distribution through 750,000 post-earthquake questionnaire responses.Source: §B.2 · p. 20
  • FloodDepth contains 150 online flood photos for estimating water depth against reference objects such as signs, vehicles, and people.Source: §B.2 · p. 20
  • Behavioral Facilitation comprises post-2018 Hokkaido earthquake data labeled with usefulness ratings based on behavioral facilitation information.Source: §B.2 · p. 20

B.3 Extraction Datasets

  • Fu et al.'s dataset uses county-level news from urban floods between 2000 and 2022 to extract disaster time and location.Source: §B.3 · p. 20
  • The geological-hazard extraction dataset contains 5,560 annotated instances, 12,980 entities, and 6,895 relations.Source: §B.3 · p. 20
  • DisasterMM pairs 6,672 relevance-classification tweets with 4,992 location-extraction tweets, whose words are tagged as location starts, continuations, or non-locations.Source: §B.3 · p. 20
  • Other extraction resources include 22,000 crisis tweets with location tags, nine-label Re´SoCIO data, and CrisisNLP subsets with rationales for classification.Source: §B.3 · p. 20

B.4 Generation Datasets

The listed generation-oriented datasets support event summarization, information-need summaries, and visual question answering about disasters and their impacts.

B.4 Generation Datasets

  • Vitiugin and Castillo's dataset uses official event reports as ground truth for generating disaster-event summaries.Source: §B.4 · p. 20
  • CrisisFACTS collects data from eight crisis events across platforms for daily stream processing and summaries tailored to information needs.Source: §B.4 · p. 20
  • DisasterQA contains disaster-related multiple-choice questions drawn from seven sources.Source: §B.4 · p. 20
  • FFD-IQA pairs 2,058 images with 22,422 question-answer ground-truth pairs about trapped-person safety and emergency-service availability.Source: §B.4 · p. 20
  • FloodNet provides 4,500 post-Hurricane Harvey question-image pairs covering counting, yes/no, and scene-condition questions about buildings, roads, and scenes.Source: §B.4 · p. 20

Related papers

Loading 2501.06932v2…