Automatic Classification of Disaster Images Based on Deep Learning

Recently, persistent catastrophic issues and advancements in science and technology have increased the need for disaster research. In this study, we propose a deep learning-based framework that distinguishes disaster images from large-scale datasets, thereby providing access to disaster-related image data. To construct an accurate dataset for our framework, disaster images were manually collected and labeled from various open datasets. Image generation and augmentation techniques were used to supplement the insufficient training dataset and enhance the classifier training of our classification framework. We built a classification framework that demonstrates over 99% accuracy in classification experiments using open datasets.