Koki Asami, Shono Fujita, Kei Hiroi, & Michinori Hatayama. (2022). Data Augmentation with Synthesized Damaged Roof Images Generated by GAN. In Rob Grace, & Hossein Baharmand (Eds.), ISCRAM 2022 Conference Proceedings – 19th International Conference on Information Systems for Crisis Response and Management (pp. 256–265). Tarbes, France.
Abstract: The lack of availability of large and diverse labeled datasets is one of the most critical issues in the use of machine learning in disaster prevention. Natural disasters are rare occurrences, which makes it difficult to collect sufficient disaster data for training machine learning models. The imbalance between disaster and non-disaster data affects the performance of machine learning algorithms. This study proposes a generative adversarial network (GAN)- based data augmentation, which generates realistic synthesized disaster data to expand the disaster dataset. The effect of the proposed augmentation was validated in the roof damage rate classification task, which improved the recall score by 11.4% on average for classes with small raw data and a high ratio of conventional augmentations such as rotation of image, and the overall recall score improved by 3.9%.
|