Classification of Soft-Story Buildings Using Deep Learning with Density Features Extracted from 3D Point Clouds

Abstract

Soft-story buildings are seismically vulnerable during earthquakes. The identification of such buildings is vital in seismic risk mitigation to assess the seismic resilience of a given urban region. Several studies have implemented deep-learning (DL) techniques to detect and classify infrastructural damage using images; however, few have focused on the detection of such buildings at the city scale. Previous models have used well-controlled imagery data instead of raw images where the targets are blocked and may thus misclassify soft-story buildings when applied to real-world data. To address this issue, this paper developed a workflow scheme that segments three-dimensional (3D) point-cloud data in a city, extracts point density features for buildings, and identifies soft-story buildings using DL models. The city of Santa Monica (California, USA) was selected as the target region for the training, validation, and testing. A naive convolutional neural network (CNN) model was proposed and compared to state-of-the-art deep CNN models trained through transfer learning (TL) techniques. The parameter sensitivity ranges for optimal performance were determined.