Abstract
Twitter can supply useful information on infrastructure impacts to the emergency managers during major disasters, but it is time consuming to filter through many irrelevant tweets. Previous studies have identified the types of messages that can be found on social media during disasters, but few solutions have been proposed to efficiently extract useful ones. We present a framework that can be applied in a timely manner to provide disaster impact information sourced from social media. The framework is tested on a well-studied and data-rich case of Hurricane Harvey. The procedures consist of filtering the raw Twitter data based on keywords, location, and tweet attributes, and then applying the latent Dirichlet allocation (LDA) to separate the tweets from the disaster affected area into categories (topics) useful to emergency managers. The LDA revealed that out of 24 topics found in the data, nine were directly related to disaster impacts—for example, outages, closures, flooded roads, and damaged infrastructure. Features such as frequent hashtags, mentions, URLs, and useful images were then extracted and analyzed. The relevant tweets, along with useful images, were correlated at the county level with flood depth, distributed disaster aid (damage), and population density. Significant correlations were found between the nine relevant topics and population density but not flood depth and damage, suggesting that more research into the suitability of social media data for disaster impacts modeling is needed. The results from this study provide baseline information for such efforts in the future.