Integration of large vision language models for efficient post-disaster damage assessment and reporting

Area of Science:

• Artificial Intelligence
• Disaster Management
• Human-Computer Interaction

Background:

• Traditional disaster response relies on human coordination, facing limitations in speed and efficiency.
• Human errors and delays in disaster response can lead to increased human and economic losses.
• Agentic Large Vision Language Models (LVLMs) present a novel approach to enhance disaster management capabilities.

Purpose of the Study:

• To introduce DisasTeller, a multi-LVLM framework for automating post-disaster management tasks.
• To explore the potential of LVLMs in improving disaster resilience and resource access, especially in underdeveloped regions.
• To assess the effectiveness of an AI-driven system in accelerating disaster response activities.

Main Methods:

• Development of DisasTeller, a framework coordinating four specialized LVLM agents powered by GPT-4.
• Implementation of automated tasks including on-site assessment, emergency alerts, resource allocation, and recovery planning.
• Evaluation of the framework's impact on coordination, information flow, and execution time in disaster scenarios.

Main Results:

• DisasTeller demonstrated the ability to accelerate disaster response activities and streamline coordination.
• The framework effectively structures information flow and automates report generation.
• Potential for significant socio-economic impact by improving resilience and resource access.

Conclusions:

• LVLM-powered frameworks like DisasTeller can significantly augment traditional disaster response methods.
• Human validation and continuous improvement of LVLM accuracy are crucial for trustworthy deployment.
• DisasTeller serves as a complementary support system, bridging the gap towards AI-driven efficiency in disaster management.