Data-efficient segmentation of collapsed building in satellite imagery using vision foundation models

The rapid and accurate identification of collapsed building following an earthquake is critical for resource allocation during the golden rescue window. However, in the early stages of a disaster, there is often a severe shortage of high-quality labeled data, making rapid deployment difficult. To investigate adaptation under limited target-domain labels, this paper proposes Satellite Image Collapsed Building Segmentation (SICBSeg), a data-efficient semantic segmentation framework for collapsed building based on visual foundation models. Built upon the Segment Anything Model (SAM), this framework introduces parameter-efficient Adapter Tuning Layers (ATL). By updating only a small number of parameters, it efficiently transfers the general visual priors of large models to complex post-disaster remote sensing scenarios. Additionally, the model incorporates learnable sparse prompt tokens, enabling end-to-end automated evaluation without human intervention. Under a data-efficient experimental setup with limited target-domain samples, SICBSeg achieves an mIoU of 0.713 and an F1 Score of 0.806. These results outperform mainstream classical convolutional networks, real-time segmentation baseline models, and improved models based on SAM. Furthermore, the framework maintains comparatively stable performance under a variety of tested image perturbations. SICBSeg performed well in practical applications in the typical disaster-stricken areas of the Yushu earthquake. It achieved precise identification of collapsed building despite the interference of complex high-altitude topography and fragmented rubble textures, demonstrating its practical potential. This study provides an engineering-oriented intelligent solution that combines high accuracy with high practical feasibility for rapid assessment of urban earthquake damage under data-scarce conditions.