Mining Social Media Data for Rapid Damage Assessment during Hurricane Matthew: Feasibility Study

Abstract

Previous research has employed social media index such as disaster-related ratio (DIRR) or damage-related ratio (DARR) and sentiment to estimate the damages during disasters. These studies mainly used predefined keywords to filter disaster- and damage-related social media data. However, many tweets containing predefined keywords (e.g., Hurricane Matthew) do not describe disaster events or their impacts. Meanwhile, previous sentiment analysis has not considered users’ tweet frequencies, which can bring in data bias. These studies also lacked a baseline to reflect disaster impacts on the public’s sentiment. Therefore, this research proposes to use supervised machine-learning approach to identify the damage-related social media data. It also analyzes users’ tweet frequencies and introduces the annual average sentiment as the baseline to calculate the normalized sentiment. Compared with previous research, the authors’ method has identified more damage-related tweets and demonstrated higher precision and recall. Correlation analysis is conducted between the social media index and the insurance claim data in Hurricane Matthew. The results show a strong and positive correlation between the DARR and claim data. A strong and negative correlation is found between sentiment and claim data. The adjusted R2 of the final regression model between damage and social media index demonstrates the feasibility of mining social media data for rapid damage assessment. The results can benefit crisis response managers in collecting real-time information and understanding timely situations during disasters.