Prof. Yu Xiao, PhD (USA)
Associate Professor Department of Landscape Architecture & Urban Planning
Hazard Reduction & Recovery Center
Texas A&M University
Landscape Architecture and Urban Planning
Local and economic development, economic resiliency, and disaster management and recovery
Bayesian Spatial Inference for Integrating Social Media and Authoritative Data for RealTime Risk Mapping
Disaster risk maps are decision support tools that can inform spatial and temporal planning for disaster mitigation, response, and recovery, which can lead to sustainable development. The existing disaster risk maps are highly dependent on low-frequency authoritative data, including hazard maps of both natural and anthropogenic threats, systems’ vulnerability and elements at risk, such as geographic and census data collected by government agencies. These maps are static in nature, not capturing the real-time changes in risk as multiple events’ hazards evolve. An gap in research is to integrate low-frequency hazard and social vulnerability measures with high-frequency real-time data, tracing the evolution of a disaster event.
In this research, we develop a framework for integrating low-frequency authoritative data, such as hazard maps generated from the U.S. Geological Survey (USGS), and socioeconomic data from U.S. Census Bureau, with high-frequency data from social media sources such as Twitter for real-time risk assessment. By combining hazard information with social vulnerability index, we assess real-time risk maps. We utilize a machine-learning algorithm to automatically detect situational awareness (SA) relevant messages and classify them into various topic categories (e.g., stock up, damage and donation) during various disaster phases. The categorization of geocoded tweets allow us to trace the spatial-temporal evolution pattern of the disaster event. Finally, we use the Bayesian inference built up on the risk mapping technique, to integrate the established prior probabilities with the real-time hazard specific information to generate posterior probabilities of risk, which allows to reduce the uncertainty of risk predictions as more evidence is being collected, and to identify sudden risk anomalies that can help to build decision-making strategies based on early warnings.
We test out the analytical framework by a case study of Hurricane Sandy that struck the northeastern states in the U.S on October 29, 2012. A total of 1,763,141 tweets posted on Twitter during October 10 and November 27, 2012 were collected and analyzed.
Bayesian risk mapping, social media data, risk assessment, risk mapping, real-time