Automation of seismic exposure models: identification of building typologies through deep learning

Abstract

Seismic exposure models characterize and quantify the population, assets, and infrastructure that could be subject to losses during an earthquake. The conventional method consists of in-person inspections, which are time-consuming and costly, lately complemented or replaced by virtual ones, which are still time-consuming. This study proposes automating the characterization of buildings (building stock), including details such as the number of stories, structural system, and construction period, by implementing a convolutional neural network (CNN) model that processes labeled images from Google Street View. To train the CNNs, the images are first filtered using an object detector to select the portion of the image containing the building of interest. Subsequently, a perspective correction is performed using a keypoints model along with a homography transformation to obtain an unskewed, orthogonal image of the façade. Results exhibit an 88.0% accuracy for structural system identification, an 80.1% mAP for the number of stories at 0.50 IoU, and a 69.4% accuracy for construction period determination. These characteristics are combined to generate a probabilistic distribution of the building taxonomy. The new models enable the identification of key buildings characteristics, providing valuable insights for risk assessment and mitigation strategies.