| Sumario: | Based on social media's potential in understanding and fortifying situational awareness, these platforms have been widely used in disaster-related research and practice. The accuracy of mining critical, actionable information from social media needs enhancement, however, to support real-time disaster management. This study demonstrates a novel framework that leverages pretrained large language models to accurately parse fine-grained, location-based information from X (formerly Twitter) for real-time disaster response. The framework categorizes and locates messages into four classes with actionable information: human requesting rescue, animal needing help, infrastructural damage, and shelter information. The performance of the framework is manually validated and proven reliable through random sampling. A comparative study is conducted with data collected from Hurricanes Harvey (2017), Irma (2017), and Ian (2022). Spatiotemporal analysis reveals that social media indexes can be an additional source for predicting damage, especially for the hurricanes that cause more indirect damage such as flood inundation than direct wind damage. Knowledge derived from social media data analysis can assist policymakers, first responders, and volunteers in supporting future disaster responses. The research proves the potential of social media and advocates for improved data access in a privacy-conscious manner, reinforcing social media's role in disaster response.
|
|---|