Share

Time-consuming building surveys may soon be replaced by automation, as shown in research led by Daniel Gomez, GEM Collaborator/ Exposure Analyst, on deep learning for exposure assessment. The study introduces an automated method for identifying building typologies, reducing the time and cost of creating building inventories for seismic risk assessment.

Automating a Manual Process

Traditionally, compiling detailed building stock data for seismic exposure assessments has relied on in-person surveys, a process that is both expensive and time-consuming. Virtual inspections through online imagery have provided some relief but continue to require considerable manual input.

The paper, Automating building typology identification for seismic risk assessment using deep learning, recognised as Editor’s Choice in the August 2025 issue of Earthquake Spectra, the journal of the Earthquake Engineering Research Institute (EERI), proposes a novel solution by applying computer vision and deep learning techniques to images from Google Street View. The methodology extracts and classifies building features such as number of storeys, structural system, and construction period (pre-code or code). 

Methodology and Accuracy

The approach employs a convolutional neural network model enhanced with pre-processing techniques. An object detector isolates building façades while a keypoint model and homography transformation correct perspective distortions, allowing the system to perform reliably even with limited data sets.

This process enables more precise classification than previous approaches, which often grouped buildings into broad categories. The study achieved an accuracy of 88% for identifying structural systems, 78% for the number of storeys, and 69% for construction period determination. These outputs were integrated into a probabilistic distribution model of building taxonomy, which can then be used to estimate seismic vulnerability.

Implications for Risk Reduction

By automating the creation of building inventories, the study addresses one of the biggest challenges in regional seismic exposure modelling – the need for reliable, detailed data at scale. The results pave the way for faster and more cost-effective seismic risk modelling in rapidly urbanising areas where large-scale in-person inspections are not feasible.

The research demonstrates how advances in artificial intelligence can strengthen the field of disaster risk reduction by providing scalable, transparent, and globally applicable tools. Its open and adaptable methodology supports broader efforts to improve seismic exposure and risk models worldwide.

Link to the study:

https://journals.sagepub.com/doi/full/10.1177/87552930251327435