ENHANCING POST-EARTHQUAKE ROAD DAMAGE ASSESSMENT USING GENERATED FAKE DATA
Journal
World Conference on Earthquake Engineering Proceedings
Journal Volume
2024
ISSN
30065933
Date Issued
2024
Author(s)
Abstract
The applications of deep learning models in post-disaster damage assessment of bridges and infrastructures using remote sensing images have raised attention in recent years. However, the lack of diverse post-disaster imagery and labeled training data is a crucial challenge for the related studies. Most existing post-disaster damage assessment models are facing an inevitable barrier resulting from the domain discrepancy. This problem is especially critical for the damage assessment of a large-scale road network because the characteristics of road systems, e.g., network pattern and material of pavement, may vary from place to place. On the other hand, the features of indirect damage caused by the surrounding built environment may also differ in different regions. This study aims to address this research gap by augmenting the training data with generated fake data. Different ways for generating fake post-disaster imagery are first discussed, including noise-based and Generative Adversarial Network (GAN) methods that mimic the damage characteristics of specific regions. The benefit and effectiveness of the proposed approach are demonstrated in the assessment of post-disaster road damage in the 2018 Sunda Strait tsunami in Indonesia. The result indicates that the domain-specific GAN-generated fake data can improve the performance of the prediction model in the target region. Findings from this study can help rapid disaster response and rescue operations planning.
Publisher
International Association for Earthquake Engineering
Type
conference paper