AI uses Google Street View to inspect buildings in hours, not months

After a massive earthquake hits, disaster response teams usually face a brutal bottleneck. Sending structural engineers door to door to inspect every building in a city can easily take months.

Chaofeng Wang, an assistant professor of AI at the University of Florida, is flipping that playbook. By feeding Google Street View and municipal street photos into deep learning models, he is turning everyday street imagery into an automated structural health monitor for entire cities.

Instead of taking months, his system wraps up an entire city's building evaluation in just a few hours.

Supercomputing meets building codes

The AI system scores buildings using guidelines from FEMA P-154, a standard established by the U.S. Federal Emergency Management Agency. By analyzing images, the model can detect wall materials, structural types, and estimated building age to assign an overall vulnerability score.

To build the training pipeline, Wang teamed up with the World Bank's Global Program for Resilient Housing to collect and annotate real-world images. Heavy lifting for inference is handled by NVIDIA DGX A100 nodes within UF’s HiPerGator supercomputer, which delivers 700 petaflops of AI power.

The output is routed to an interactive map portal, showing city officials the safety scores, structural categories, and exterior materials for individual properties.

So far, the model has been tested across roughly eight cities internationally, hitting over 85% accuracy on FEMA P-154 scoring standards. Local authorities in Mexico are actively testing the platform, and organizations across parts of Africa, Asia, and South America have registered interest.

Global paint and dense alleys

Training a model on one region’s buildings rarely translates neatly across borders. Wang noted that regional architectural quirks and urban density variations remain the biggest hurdles.

Something as mundane as exterior paint color can trick the computer vision model into misidentifying brick, concrete, or wood. If the system has not seen a specific local vernacular, predictions can wobble.

Freshness is another hurdle. Google Street View updates every few years, which works fine for long-term disaster preparation and baseline risk mapping, but fails in off-grid rural areas or immediately after a disaster. In those scenarios, local governments must capture and upload their own photos.

Wang’s team is now building application programming interfaces and a user-friendly app so agencies can plug the analysis directly into their existing workflows.

Why it matters

Resilience planning has always suffered from high costs and structural engineer shortages, leaving vulnerable communities exposed. Wang’s approach is clever because it repurposes a free, ubiquitous data layer—consumer street mapping—into high-value municipal safety intelligence.

For local governments, it shifts the strategy from reactive disaster recovery to proactive retrofitting, pinpointing weak structures before an earthquake strikes. It could eventually serve property insurers and urban planners looking for low-cost portfolio exposure audits.

That said, we think street-level AI has clear boundaries. Exterior pixels cannot reveal internal framing decay, illegal load-bearing alterations, or shifting foundations. For now, it works best as a triage scanner to help officials decide where to dispatch human inspectors first, rather than a total replacement for boots on the ground.