UCLA

Seismically vulnerable, especially collapse-prone, buildings often represent the greatest life-safety hazard worldwide. Identifying these buildings is the first step in seismic risk mitigation efforts for a given urban region’s resilience. This thesis aims to devise a workflow for the application of state-of-the-art deep neural networks (DNNs) for detecting and classifying seismically vulnerable buildings using three-dimensional point clouds. A number of prior studies have focused on using 2D image data in the field of structural health monitoring for damage recognition. The performance of those approaches for building classification at regional scales is highly dependent on well-controlled imagery data and may not be guaranteed when applied to real-world data. The present study, therefore, differs from prior studies in that it uses 3D point clouds, which implicitly contain depth information that can become a highly useful feature for training DNNs. Here, a specific DNN, namely, PointNet, is used for detecting soft-story buildings, which are ubiquitous in west-coast cities of the US. A workflow is devised for binary classification of point clouds into soft-story and non-soft-story points as well as the segmentation and association of classified point clouds with specific addresses. Over 10 billion point clouds obtained from the city of Santa Monica in California are manually labeled and split into training, validation, and testing sets, and the sensitive ranges of DNN hyperparameters are investigated to obtain the good performance.