Item

Abstract

Artificial intelligence, specifically deep learning (DL), has been increasingly used in remote sensing technology for rapid disaster assessment. Despite having high accuracy, the approach requires numerous samples to maintain its performance. However, in an emergency response phase, training samples are often unavailable. Moreover, generating ground truth data is laborious and time-consuming, which impedes the fulfillment of damage information for emergency response needs. Given this background, we utilized historical disaster events collected worldwide to recognize damage in new unseen locations. We compared the performance of multiple DL convolutional neural networks (e.g, ResNet34, ResNex, and SwinTransformer). We used the trained model to estimate the damage in the 2011 Tohoku Tsunami to illustrate the usefulness of both historical data and DL models for mapping a new disaster event. Preliminary results show that SwinTransformer outweighs the performance of others CNN-based models. The accuracy of the transformer-based model reaches 85% compared to 82% for both ResNet34 and ResNext. However, when the best performance model is tested on the 2011 Tohoku Tsunami, the performance drops to 23% accuracy. This reduction in accuracy may be attributed to the difference in sensing conditions and data distribution between the global and the Tohoku datasets.