Deep Learning-Based Spatial Analytics for Disaster-Related Tweets: An Experimental Study

Abstract

Online social networks are being widely used during unexpected large-scale disasters not only for sharing latest news but also requesting emergency rescues. Particularly, social network posts with their location information are becoming more important because they can be utilized for emergency management, urban planning, and various studies to understand effects of the disasters. Despite their importance, the percentage of such posts is generally tiny. In this paper, to address the location sparsity problem on Twitter in the event of disasters, we propose a deep learning-based framework to spatially analyze the disaster-related tweets by focusing on classifying tweets from affected areas of disasters. We also study effects of different deep learning architectures and input embedding techniques for this classification task. Our experimental results demonstrate that our ConvNet model with the Word2vec word embedding has the highest classification accuracy.