Using Machine Learning, Social Media Images, and Journalists to Improve Disaster Resilience and Response

Abstract

Purpose: We explored the feasibility of generating a real-time, priority-driven map of infrastructure damage during a natural disaster. We examined the utility of harnessing Twitter data during/after a natural disaster to optimize response and rapidly assess infrastructure damage across a region. Design/Methodology/Approach: We strategically selected journalist networks to identify sources generating valuable infrastructure-damage data. Hurricane Florence-related tweets were collected using the REST Twitter API with Python, from September 13-18, 2018. Tweets were classified by source (news or citizens). We utilized Google AutoML Vision software to develop the machine learning image classification model; 80% was used for training; 10% was used for validation, and 10% was used for testing. Findings: The average precision of the model was 90.6%, and the average recall was 77.2%. The F1 score for this model is 83.4 (see Figures 10, 11, 12, and 13). Research Limitations/Implications. Data were analyzed post-hoc; a logical next step is to test this model in real-time. Our sample contained 11,638 images, a more expansive image dataset would likely yield greater precision, recall percentages, and F1 scores. In the future, establishing strategic networks of journalists ahead of disasters will reduce the time it takes to identify disaster-response targets to focus relief and recovery efforts in real-time--saving lives and mitigating harm.Originality/ValueBy using this classification system, we demonstrate strategies to create a nationwide network of journalists and citizen scientists that contribute information about post-disaster infrastructure damage and identify methods that can be leveraged in real-time to focus first-response, disaster, and recovery efforts.

Keywords: hurricanes, Twitter, Computer vision, Machine learning, Disasters, Journalism