Rapid post-earthquake building damage mapping with object-oriented SAM fine-tuning: a case study of the 2023 Jishishan earthquake

Earthquakes pose severe risks to lives and infrastructure, demanding rapid and accurate building-damage assessments to guide response. High-spatial-resolution (HSR) imagery is essential for this task, yet conventional convolutional neural networks (CNNs) can be limited in capturing long-range dependencies, especially in dense urban fabrics. We introduce S2AC-Net, an operational framework that fuses global context with local features for post-earthquake damage mapping. In S2AC-Net, the Segment Anything Model (SAM) replaces traditional multi-resolution segmentation for pre-event imagery: a frozen Vision Transformer encoder with lightweight adapters produces building-probability maps, which are merged with SAM masks to delineate building footprints. Spectral and texture features from pre- and post-event images are then mapped to these objects to form feature vectors, which—together with field survey labels—drive a CNN classifier for multi-class damage grading. Applied to the December 2023 Jishishan earthquake (Gansu, China), S2AC-Net achieved 0.928 overall accuracy for building localization and an overall F1 of 0.882 for damage assessment. These results demonstrate that integrating SAM-based global context with object-level feature fusion overcomes key limitations of purely CNN-based pipelines and provides a scalable pathway for rapid, reliable remote-sensing–based disaster assessment and urban resilience planning.