Improving the Assessment of Post-Earthquake Building Damage in Underdeveloped Regions with Vision Transformers

Assessing building damage following an earthquake is vital for first responders to effectively target their efforts in disaster-stricken areas. Satellite imagery is a powerful tool for visualizing such damage, particularly in underdeveloped regions where infrastructure is limited and access to disaster sites is challenging. This paper focuses on enhancing the process of classifying building damage following an earthquake by utilizing high-resolution satellite imagery and the state-of-the-art Vision Transformers (ViT) model. Experiments are carried out using two real-world datasets from the Ludian and Yushu earthquakes, contrasting the effectiveness of ViTs with sophisticated CNNs like ResNet50, Inception-V3, and EfficientNet-B0. The results show that ViT can more effectively assess rural buildings compared to other models. It also demonstrated better generalization across different earthquake scenarios and stays robust when trained and tested on smaller datasets. Furthermore, we proposed a new architecture that incorporates a CNN-based Inception module so the local features on the damaged buildings can be better extracted. improved the models ability. The results show that ViTs have the potential to be a reliable and powerful tool for constructing damage assessments in disaster-affected areas, providing a more precise and effective solution than conventional CNN-based techniques. Furthermore, the proposed Inception-enhanced ViT architecture presents a viable path for further study, with the potential to be refined and validated on larger datasets and applied to a broader range of disaster scenarios