Deep Domain Adaptation Approach for Classification of Disaster Images

Abstract

In the last decade, emergency responders, government organizations and disaster response agencies have certainly acknowledged that microblogging platforms such as Twitter and Instagram can be an important source of actionable information at the time of disaster. However, researchers feel that analyzing social media data for disasters using supervised machine learning algorithms is still a challenge. During the first few hours of any disaster when labeled data is not available, the learning process gets much delayed. The first objective of this study is to explore the domain adaptation techniques by making use of labeled data of some previous disaster along with the abundance of unlabeled data that is available for the current disaster. The second objective is to apply these domain adaptation techniques on disaster-related imagery data from the microblogging platforms since imagery data has largely been unexplored as compared to textual content. To achieve these objectives, domain adaptation methods would be applied to classify the images of an ongoing disaster as informative versus non-informative. This study, for the first time, proposes a semi-supervised domain adaptation technique where the classifier is trained on three types of data, labeled data of the previous disaster, unlabeled data of current disaster and a small batch of labeled data of current disaster. Our experiments have been performed on Twitter images corresponding to three disasters. The experimental results show that there is an improvement in the accuracy of the classification model if a small batch of labeled target images is also added along with the unlabeled target images and labeled source images at the time of training. The experiment aims to make the best use of the labeled data of a previous disaster to analyze the current disaster without any delay for better response and recovery.