Earthquakes and other natural disasters rank among the most destructive events, causing widespread loss of life and severe economic consequences globally. A primary consequence of earthquakes is the large-scale collapse and damage of buildings. The rapid advancement of high-resolution remote sensing technology, offering extensive coverage and multi-temporal capabilities, combined with deep learning methods, has opened new possibilities for accurately and efficiently detecting and assessing building damage to support crisis management. However, pre- and post-disaster images are often acquired under varying temporal, lighting, and weather conditions, complicating the task of accurately identifying building damage levels. This study proposes a Siamese network based on UNet to address these challenges, enabling the assessment of building damage using satellite imagery following earthquakes. The network leverages multi-scale feature differentiation to model spatial and temporal semantic relationships, addressing the issue of intra-class semantic variation. The proposed method was evaluated on the xBD disaster damage dataset and the 2023 Morocco earthquake dataset, achieving an overall accuracy of 95.5% and a kappa coefficient of 76.0%. These results highlight the potential of AI-driven solutions to meet the critical demands for speed and accuracy in disaster response scenarios.