[HTT14-08] A Comparative Analysis of the Impacts of Various Natural Disasters on Human Mobility based on Geotagged Photos
★Invited Papers
Keywords:Natural disasters, Human mobility, Geotagged photos, Flickr, Convolutional Neural Networks
Natural disasters perturb the daily movement of individuals. Researchers have previously investigated human mobility patterns under steady and unsteady conditions using text-based geotagged data collected from social media giants of the likes of Twitter and Weibo characterized by their low cost and availability. However, with a growing number of new limitations to access to their historical databases, new alternatives of geodata sources of mass have been proposed. Image sharing platform Flickr is one of the suggested options given not only its free-to-use Application Programming Interface (API) but also the type of information that can be extracted from photos using advanced techniques of deep learning. This might give an insightful output regarding human mobility patterns in urban areas during steady conditions as well as during temporary unstable states caused by different types of disasters. In the light of this, the aim of the present study is to analyze and compare the impacts of several past natural disasters (earthquake, typhoons, etc.) that occurred in the Tokyo metropolitan area (TMA) on human mobility patterns based on Flickr geotagged data during three periods: before, during, and after each disaster. First, out of 22 natural catastrophic events that occurred between 2008 and 2019, we chose the significant ones based on the number of Flickr active users and the availability of images that were taken during that period. Second, based on occurrence day(s) and relevant weather information associated with each type of disaster, we selected steady and perturbation periods (before, during, after) for each disaster. Third, we analyzed mobility patterns through multiples indicators namely displacement, radius of gyration, and mean square displacement. We then compared the obtained results of all disasters. Preliminary results show that the probability distribution of displacement and radius of gyration follow a truncated power-law, which is in agreement with previous studies. Next, we developed a transfer learning-based convolutional neural network (CNN) model to classify images into indoor and outdoor categories. This classification helped to reconstruct popular trajectories by type of environment and duration of stay under steady and unsteady conditions. The results of this study are expected to stimulate future research on the impacts of natural disasters on the mobility of foreign tourists given the popularity of Tokyo as a touristic hotspot and yet as a metropolitan area of recurring disasters with scarce spatial data and studies on tourists mobility patterns.