Abstract Earthquakes present a significant threat to human life and infrastructure, highlighting the urgent need for rapid and accurate building damage assessments to inform disaster response efforts. High Spatial Resolution (HSR) remote sensing images, which offer detailed surface information, are indispensable for such assessments. However, existing methods, especially those based on Convolutional Neural Networks (CNNs), frequently fail to capture global correlations between local image patches, reducing their effectiveness in densely populated urban areas. This gap underscores the need for a more robust approach that integrates global context with local features. Here, we present S2AC-Net, a novel framework for post-earthquake building damage assessment. In this framework, the Segment Anything Model (SAM) replaces traditional multi-resolution segmentation in processing pre-disaster images by utilizing frozen image encoding parameters and adapters within a Vision Transformer structure to predict building probabilities. Building delineation is achieved by merging segmentation results with probability predictions, and spectral and texture features from pre- and post-disaster images are mapped onto building regions to construct feature vectors. These vectors, when combined with field survey data, enable damage level assessment using CNNs. When applied to the December 2023 Jishishan earthquake, S2AC-Net achieved a building localization accuracy of 0.928 and an overall accuracy of 0.882. These results demonstrate S2AC-Net's effectiveness in overcoming the limitations of traditional CNNs, advancing the field of remote sensing-based disaster assessment, and providing a scalable solution for urban resilience planning.
Yang et al. (Mon,) studied this question.