Instance segmentation techniques for seismic building structural type estimation from remote sensing imagery – Evidence from Xi'an city, China

Collecting exposure information for seismic risk assessment if frequently a labor-intensive and costly aspect. This study reveals the potential of automatically determining Seismic Building Structure Types (SBSTs) utilizing remote sensing imagery and instance segmentation models. A comprehensive process is introduced, which encompasses (i) data acquisition from remote sensing imagery; (ii) compilation of training data for subsequent supervised model learning, including clipping, resizing, zero-padding, labeling, and augmentation; (iii) and supervised model learning using the YOLO Series. Regarding the latter, we implement a set of seventeen pretrained models from YOLOv5, v7, v8 and v11 and provide an exhaustive experimental evaluation. The ancient Xi'an city wall is employed as the research area to evaluate the models' classification accuracy based on the buildings within it. The findings are as follows: A relatively larger model size and better adaptability of the model to the task lead to better performance in instance segmentation, allowing YOLOv7x-seg to outperform other models. By comparison, the mean average precision value for singular tasks, such as height and material instance segmentation, surpasses that of the comprehensive task, i.e., SBST instance segmentation, with effectiveness increasing from SBST, to height, and finally, to material. From an application standpoint, the models effectively identify buildings across various urban layouts, including buildings in open scenes, regularly arranged structures, and dense, irregularly arranged developments. However, the models still occasionally exhibit instances of missed detections or false positives. Nevertheless, our work underlines the great potential for a rapid assessment of crucial seismic exposure properties in complex built environments.