Kuo W.N.Tsai F.-J.SZU-YUN LIN2026-03-162026-03-162024https://www.scopus.com/record/display.uri?eid=2-s2.0-105027867344&origin=resultslisthttps://scholars.lib.ntu.edu.tw/handle/123456789/736381After a severe earthquake, rapid damage assessment is critical for disaster response and rescue operations. In recent years, the application of deep learning and image processing technologies in the identification of disaster damage is increasing. This study focuses on the large-scale post-earthquake building damage assessment combining parameter analysis and an image segmentation model. First, a Siamese neural network is used to predict the location and damage level of the buildings, adopting pre-disaster and post-disaster remote sensing image pairs as the inputs. The images in the xBD dataset and the satellite images of the historical events, such as the 2010 Haiti earthquake and the 2023 Turkey earthquake, have been collected and labeled as training data. On the other hand, a gradient-boosting regression model is trained for damage level classification, adopting basic building information, e.g., building height, construction year, occupancy, structural materials, etc., and ground motion intensity parameters as the inputs. For this parametric-based machine learning model, the training data are collected from the post-disaster building damage reports and the ShakeMap record of the earthquakes, including the 2015 Nepal earthquake, the 2016 Meinong earthquake, and the 2018 Hualien Earthquake in Taiwan. A Multi-modal Learning framework integrating the above models is proposed to enhance post-disaster large-scale post-earthquake building damage assessment. The results show that the integrative Multi-modal model improves the prediction accuracy compared to the ones by merely image segmentation deep learning model or parametric-based machine learning model.falseconference paper2-s2.0-105027867344