Automating building typology identification for seismic risk assessment using deep learning

Driven by rapid urbanization and heightened seismic risk concerns, efficient methods for developing regional seismic exposure assessments are advantageous. By leveraging deep learning and computer vision techniques, this study presents a novel approach for automating the identification of building typologies. The detailed building stock required for seismic exposure assessment has been traditionally achieved through time-consuming and costly in-person inspections. Recently, virtual inspections have emerged as a more efficient alternative, but they still require significant manual effort. This study proposes a methodology for automating the characterization of buildings, including details such as the number of stories, structural system, and construction period (pre-code or code), by implementing a convolutional neural network model that processes labeled images from Google Street View. A key innovation of this study is the integration of pre-processing techniques, including an object detector to isolate building façades and perspective correction using a keypoint model and homography transformation, enabling robust performance even with a small data set. This research advances prior methods by classifying individual stories rather than grouping them into broad taxonomic ranges, providing greater precision and applicability for seismic exposure modeling. The results show an 88% accuracy for structural system identification, a 78% accuracy for the number of stories, and a 69% accuracy for construction period determination. These characteristics are integrated into a probabilistic distribution model of building taxonomy that informs about their potential seismic vulnerability. The proposed procedures streamline the development of building stock and seismic exposure models, thus facilitating their use for seismic risk modeling at a regional scale.