| description abstract | The effectiveness of infrastructure resilience relies on the seamless extraction of information, timely acquisition of critical knowledge, and heightened situational awareness. The ongoing utilization of digital citizen communication through social media with response organizations during disasters remains a valuable avenue for disseminating information, ensuring the effective utilization of public resources in emergency response to crisis events. Public agencies can use this information to examine community sentiments and discussions to assess, determine, and prioritize critical areas in need of assistance. However, there are limitations on harnessing precise geolocation information from social media, as well as a lack of mitigating bias of machine learning models used during such events. These limitations can restrict emergency management personnel’s ability to locate and promptly delineate actionable insights. Here, we propose a semisupervised machine learning model that utilizes approaches such as transfer learning, topic modeling (i.e., Latent Dirichlet Allocation), and natural language processing to augment data from historical and current social media posts (i.e., Twitter) with community-driven application alerts (i.e., Waze) to achieve further evidence on the location and context of emergency events. The model is designed to also mitigate machine learning bias using the Wells–Du Bois protocol. A framework was developed for this process and is illustrated through a case study on Hurricane Ian and three previous hurricanes that occurred in Florida. This fusion provides increased situational awareness and may enhance the speed of emergency response. This study establishes a foundation for equitable, real-time crisis event detection, expanding organizations’ response capacity in allocating resources and reducing harmful effects of disaster, particularly within public infrastructure systems. | |
