Multi-temporal deep learning-based social media analysis for disaster relief

Abstract

In recent years, using social media images to assess natural disasters and the structural damage they cause has become an increasingly prevalent research approach. Particularly, deep learning methods have enabled rapid assessment of crisis situations on the ground. As social media platforms contain firsthand accounts of these devastating instances during and after events, they provide an unprecedented opportunity for disaster relief efforts. Comparatively, for example, while satellite imagery has been previously lauded for its effective use in training deep learning models, concerns about cost of obtainment and ease of access make social media, with its multitemporal properties, potentially more useful. These computational methods provide for rapid and efficient allocation of resources and personnel, saving lives and property and minimizing economic loss. In this position paper, we discuss how social media imagery, as extracted in real time from a variety of networks, can be the most useful source of data for disaster relief when combined with machine learning and computer vision techniques, enabling effective deployment pipelines in mobile applications for use by individuals, NGOs, and local governments in disaster-prone areas.