DisImpact Framework Uses MLLMs to Quantify Disaster Impacts from Social Media

A new research paper presents DisImpact, a framework that leverages Multimodal Large Language Models (MLLMs) to systematically quantify the physi-social impacts of natural disasters through social media analysis. Traditional surveys and reports struggle to capture the cascading social consequences of disasters, but DisImpact taps into real-time, multimodal content from platforms like Twitter and Facebook. The framework first classifies posts into ten distinct impact categories spanning both physical damage (e.g., infrastructure destruction) and social effects (e.g., displacement, mental health). It then builds a weekly disaster impact index that combines the relative prominence of each category with public engagement intensity, enabling direct comparison across categories and aggregation into higher-level patterns.

The researchers validated DisImpact using authoritative ground-truth data from FEMA Public Assistance records and NASA FIRMS fire detections, finding consistent lead-lag correlations that confirm the framework's accuracy. Temporal and spatial analyses reveal that physical impacts tend to peak during the disaster and remain localized to directly affected regions, while social impacts often emerge later and spread more broadly across time and geography. This is the first comprehensive framework to quantify both physical and social disaster dimensions using multimodal social media data, offering a scalable tool for emergency responders and policymakers to assess real-time impacts and inform resource allocation.