Detecting and Geolocating City-Scale Soft-Story Buildings by Deep Machine Learning for Urban Seismic Resilience

Abstract

Seismic resilience is of great concern and vital importance for cities in earthquake zones. It is not only desirable but also mandatory for the cities to prepare an emergency response plan for possible seismic events. One important task is to identify the seismic vulnerable buildings, e.g., soft-story (SS) buildings. The identified SS buildings can be retrofitted to minimize the risk of possible damage and the mitigation plan can be made accordingly to allocate the needed resources to the vulnerable buildings when earthquake strikes. However, it is time-consuming and costly for structural engineers to identify SS buildings manually by walking through each street. This paper presents an integrated approach for automatically detecting and geolocating SS buildings at a city scale. The approach proceeds in multiple steps, including (1) obtaining a list of addresses of engineer-identified SS buildings in city of Santa Monica, CA; (2) extracting the Google Street View images of the SS buildings; (3) labeling the SS building in the images; (4) training a deep convolutional neural network with the annotated images; and (5) testing the trained model on an independent image data set. The detected SS buildings are geocoded in Google map for users to verify the results quickly and virtually.