Addressing the inefficiency of traditional post-disaster response relying on manual collaboration, a study proposes DisasTeller, a multi-agent framework powered by Large Vision Language Models (LVLMs) with GPT-4o as its core. Coordinating four specialized agents (expert, alert, emergency, and assignment teams) and integrating tools like image interpretation and map annotation, the framework automates key tasks including post-disaster damage grading, emergency alert issuance, resource allocation, and reconstruction planning. Validated by the Wajima City earthquake case in Japan, DisasTeller completes all tasks within minutes, with GPT-4o demonstrating significantly higher assessment accuracy than Gemma3. The research also identifies limitations such as error propagation in early-stage assessments, emphasizing the necessity of human-AI collaboration to ensure reliable real-world deployment.
Traditional post-disaster response heavily relies on human teamwork, which is constrained by human limitations such as fatigue and slow mobility, leading to delayed critical actions and increased human and economic losses. With the global economic losses from natural disasters reaching US$250 billion in 2023, there is an urgent need for efficient response strategies. To address this challenge, the research introduces DisasTeller, a multi-Large Vision Language Model (LVLM) framework centered on GPT-4. This framework coordinates four specialized agents to automate core post-disaster tasks, including on-site damage assessment, emergency alert issuance, resource allocation, and recovery planning, aiming to streamline information flow and reduce response time, especially benefiting underdeveloped regions with limited infrastructure.
DisasTeller’s modular design consists of four functional agents and four supporting tools, enabling a structured workflow that mimics real-world disaster response teams. The expert team agent first analyzes post-disaster images and uses the European Macroseismic Scale (EMS-98) for damage grading (G1 to G5). It then shares results with the alerts, emergency, and assignment team agents. Equipped with tools like ImageInterpretationTool, WebSearchTool, FileReadTool, and MapAnnotationTool, the agents collaborate to generate alert maps, emergency reports, human resource allocation plans, public notices, and reconstruction blueprints. The framework also supports user-friendly interfaces such as mobile apps, allowing non-experts to submit real-time images and access localized guidance, thus enhancing overall responsiveness.
To validate DisasTeller’s effectiveness, the research conducted a case study on the earthquake in Wajima City, Japan, along with additional tests on flood and bushfire scenarios. The results demonstrated that the framework could autonomously complete damage grading, map annotation, and report generation within minutes, a significant reduction compared to the weeks-long human-led process. A performance comparison between GPT-4o and Gemma3-27B showed that GPT-4o outperformed the latter in both intermediate tasks and final report quality, particularly in technical, structured outputs like reconstruction plans and resource allocation reports. Evaluations by both GPT-4o and civil engineering experts confirmed the coherence and actionability of DisasTeller’s outputs.
Despite its promising performance, DisasTeller faces several limitations that require careful mitigation for real-world deployment. A key concern is the risk of error propagation, where inaccuracies in early-stage damage grading can affect downstream decisions such as resource allocation. The framework also struggles with spatial reasoning and geospatial integration, as seen in occasional shelter suggestions near collapsed structures. Additionally, data acquisition challenges in inaccessible disaster zones and the limited interpretability of closed-source LVLMs pose barriers to trustworthiness. To address these issues, the research proposes solutions like human-in-the-loop validation, integration with GIS platforms, and adoption of open-source models to enhance transparency and traceability.
In conclusion, DisasTeller represents a significant advancement in AI-assisted post-disaster management, showcasing the potential of multi-LVLM agent systems to accelerate response workflows and reduce regional disparities in disaster resilience. While the framework cannot replace human experts, it serves as a valuable complementary tool that standardizes assessment processes and provides rapid, actionable insights. Future research should focus on improving LVLM accuracy in domain-specific tasks, integrating real-time multimodal sensor data, and strengthening human-AI collaboration mechanisms. Continued partnership between AI developers, disaster management experts, and policymakers will be essential to realize the safe and effective deployment of such systems in real-world disaster scenarios.
Sources:
Nature Communications
https://www.nature.com/articles/s41467-025-68216-z .
Provided by the IKCEST Disaster Risk Reduction Knowledge Service System
Comment list ( 0 )
