Researchers from the National Science Foundation’s
SimCenter, an engineering community focused on
modeling the impact of natural hazards, built the
Building Recognition using AI at Large-Scale suite
of tools. BRAILS creates enhanced building databases
for cities by running artificial intelligence-powered
simulations on high-performance computers at the Texas
Advanced Computing Center (TACC) at the University
of Texas at Austin.
Scientists are using data from Google Maps and satellite
images to power artificial intelligence applications that
can automatically identify characteristics of city’s buildings
that would be vulnerable in an earthquake, hurricane or tsunami.
Researchers from the National Science Foundation’s SimCenter,
an engineering community focused on modeling the impact of
natural hazards, built the Building Recognition using AI at
Large-Scale suite of tools. BRAILS creates enhanced building
databases for cities by running artificial intelligence-powered
simulations on high-performance computers at the Texas Advanced
Computing Center (TACC) at the University of Texas at Austin.
"We want to simulate the impact of hazards on all of the buildings
in a region, but we don't have a description of the building
attributes," said Charles Wang, a postdoctoral researcher at
the University of California, Berkeley, and the lead developer
of BRAILS told TACC News. "Using AI, we are able to get the
needed information. We can train neural network models to infer
building information from images and other sources of data."
The basic BRAILS framework uses computer vision to automatically
pull building details – architectural features such as roofs,
windows, and chimneys -- from satellite and ground-level images
found in Google Maps and merges these with other datasets,
such as Microsoft Footprint Data and OpenStreetMap. BRAILS users
can also enhance the data with tax records, city surveys and
other information to get more accurate assessments.
A crowdsourcing effort has contributed some of the data labeling.
Volunteer with the SimCenter’s Building Detective for Disaster
Preparedness project identified specific architectural features
of structures, such as garages, roofs and adjacent buildings,
that are then used to train additional machine learning modules.
The citizen-science project was launched on March. Within a
few weeks, a thousand volunteers had annotated 20,000 images, Wang said.
The researchers ran a series of testbeds to determine the
accuracy of the AI-based models. Each generated an inventory
of a city’s structures and simulated the impact of a hazard based
on historical or plausible events. The team has created
testbeds for earthquakes in San Francisco, and for hurricanes
in Lake Charles, La., the Texas coast and Atlantic City, N.J.
To train the BRAILS modules and run the simulations, the researchers
relied on TACC’s supercomputers -- notably Frontera, the fastest
academic supercomputer in the world, and Maverick 2, a
GPU-based system designed for deep learning.
"The hazard event simulations -- applying wind fields or ground
shaking to thousands or millions of buildings to assess the impact
of a hurricane or earthquake -- requires a lot of computing resources
and time," Wang said. "For one city-wide simulation, depending
on the size, it typically takes hours to run on TACC."
When asked about the performance of BRAILS, Wang touted the accuracy.
"For some models, like occupancy, we are seeing the accuracy is
close to 100%,” he said. “For other modules, like roof type,
we're seeing 90% accuracy."
The researhers outlined the framework in the February 2021 issue
of Automation in Construction and presented a testbed for Hurricane
Laura the 2020 hurricane that made landfall in Louisiana at the
2021 Workshop on SHared Operational REsearch Logistics In the
"Our objectives are two-fold," Wang said. "First, to mitigate the
damage in the future by doing simulations and providing results
to decision- and policy-makers. And second, to use this data to
quickly simulate a real scenario – instantly following a new event,
before the reconnaissance team is deployed. We hope near-real-time
simulation results can help guide emergency response with greater
Provided by the IKCEST Disaster Risk Reduction Knowledge Service System