Seeing the Storm: Leveraging Multimodal LLMs for Disaster Social Media Video Filtering

Abstract

Emergency management relies on the rapid triage of information to respond appropriately to disaster events. Social media platforms can provide emergency managers with ground-level insights, and videos, in particular, offer an immersive medium for understanding public responses and on-the-ground conditions. However, the overwhelming volume of irrelevant or redundant videos complicates their use for emergency response. This paper investigates the use of multimodal large language models (MLLMs)–specifically the Gemini 1.5 flash model–to automate the identification of relevant videos shared on X (formerly Twitter) during hurricanes. We develop and evaluate a framework to test the accuracy of different prompting styles and question strategies. By identifying the most effective prompting techniques, this study lays the groundwork for a systematic approach to filtering social media videos, enabling emergency managers to focus on the most pertinent content and make timely, informed decisions.