REAL-TIME CLASSIFICATION OF DISASTER IMAGES FROM SOCIAL MEDIA WITH A SELF-SUPERVISED LEARNING FRAMEWORK

Many studies have shown that disaster images posted on social media by users can be helpful to establish situational awareness and support disaster management. As such, it is important to develop a real-time framework to automatically classify disaster related images on social media from not informative ones. Unfortunately, most of the existing methods are based on supervised learning which requires manual data labeling. Meanwhile, some studies adopt unsupervised learning classification, which avoids manual labeling, but results in lower accuracy. To fill the research gap, this paper built upon MoCo model and developed a self-supervised learning classification framework. Using social media images of seven real disaster events as case studies, the results demonstrate that the proposed method without the manual labeling, can reach relatively high accuracy for disaster image classification by comparing with the state-of-the-art supervised learning and unsupervised learning models.