Image Classification using Pre-trained Deep Network for Density-based Spatiotemporal Analysis System

Abstract

Recently, during natural disasters like, earthquakes, typhoons, flood, and heavy snowfall, people actively post messages that mention situations people are facing through social media sites. Therefore, the enhancement of situation awareness in the real world using social data is one of the most attractive research topics. In our previous work, we developed a density-based spatiotemporal analysis system to identify topic-related areas in which there are a huge number of geo-tagged tweets related to a topic are posted. In this paper, we propose a novel density-based spatiotemporal analysis system with a photo image classifier using the pre-trained deep network in order to enhance situation awareness by showing accurate topic-related photos. The photo image classifier using a support vector machine (SVM) based on the pre-trained deep network is integrated into the conventional density-based spatiotemporal analysis system. To evaluate the proposed system, we used actual tweet data sets related to weather topics, "heavy rain" and "heavy snow," in Japan. The experimental results showed that the proposed system can classify photo images related to these weather topics more sensitively compared with our previous method.