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

This study explores the potential of agentic Large Vision Language Models (LVLMs) to enhance the speed and efficiency of post-disaster response, addressing human limitations that can delay critical actions and increase losses. It introduces DisasTeller, a multi-LVLM framework designed to automate key post-disaster management tasks, including on-site damage assessment, emergency alerts, resource allocation, and recovery planning. By coordinating four specialised LVLM agents with GPT-4 as the central controller, the system accelerates information processing, improves coordination, and reduces human execution time. Evaluation results demonstrate notable gains in workflow efficiency and report generation, while also revealing risks related to error propagation from early-stage assessments. The study emphasises the continued need for human oversight and accuracy improvements, positioning DisasTeller as a complementary decision-support system that bridges traditional disaster response practices with emerging AI-driven capabilities, while underscoring the importance of safe and trustworthy deployment.